From 1936e4bdfd776f78f9fe44f77ce66066fd166360 Mon Sep 17 00:00:00 2001 From: Ian Romanick Date: Mon, 17 Mar 2008 15:45:52 -0700 Subject: cell: Initial code-gen for alpha / stencil / depth testing Alpha test is currently broken because all per-fragment testing occurs before alpha is calculated. Stencil test is currently broken because the Z-clear code asserts if there is a stencil buffer. --- src/gallium/drivers/cell/spu/spu_per_fragment_op.h | 32 ++++++++++++++++++++++ 1 file changed, 32 insertions(+) create mode 100644 src/gallium/drivers/cell/spu/spu_per_fragment_op.h (limited to 'src/gallium/drivers/cell/spu/spu_per_fragment_op.h') diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h new file mode 100644 index 0000000000..6571258699 --- /dev/null +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h @@ -0,0 +1,32 @@ +/* + * (C) Copyright IBM Corporation 2008 + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * on the rights to use, copy, modify, merge, publish, distribute, sub + * license, and/or sell copies of the Software, and to permit persons to whom + * the Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL + * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, + * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR + * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE + * USE OR OTHER DEALINGS IN THE SOFTWARE. + */ + +#ifndef SPU_PER_FRAGMENT_OP +#define SPU_PER_FRAGMENT_OP + +extern qword +spu_do_depth_stencil(int x, int y, qword frag_mask, qword frag_depth, + qword frag_alpha, qword facing); + +#endif /* SPU_PER_FRAGMENT_OP */ -- cgit v1.2.3 From 284ab5a6127f8b452acaa0e10ac1d9ebc87fac3e Mon Sep 17 00:00:00 2001 From: Brian Paul Date: Wed, 10 Sep 2008 18:22:00 -0600 Subject: cell: checkpoint commit of new per-fragment processing Do code generation for alpha test, z test, stencil, blend, colormask and framebuffer/tile read/write as a single code block. Ian's previous blend/z/stencil test code is still there but mostly disabled and will be removed soon. --- src/gallium/drivers/cell/common.h | 20 +- src/gallium/drivers/cell/ppu/Makefile | 1 + src/gallium/drivers/cell/ppu/cell_gen_fragment.c | 530 +++++++++++++++++++++ src/gallium/drivers/cell/ppu/cell_gen_fragment.h | 38 ++ src/gallium/drivers/cell/ppu/cell_state_emit.c | 31 +- .../drivers/cell/ppu/cell_state_per_fragment.c | 2 +- src/gallium/drivers/cell/spu/Makefile | 2 +- src/gallium/drivers/cell/spu/spu_main.c | 53 ++- src/gallium/drivers/cell/spu/spu_main.h | 23 + src/gallium/drivers/cell/spu/spu_per_fragment_op.c | 231 ++++++++- src/gallium/drivers/cell/spu/spu_per_fragment_op.h | 11 + src/gallium/drivers/cell/spu/spu_tri.c | 30 ++ src/gallium/winsys/xlib/xm_api.c | 7 +- src/gallium/winsys/xlib/xm_winsys.c | 35 ++ 14 files changed, 998 insertions(+), 16 deletions(-) create mode 100644 src/gallium/drivers/cell/ppu/cell_gen_fragment.c create mode 100644 src/gallium/drivers/cell/ppu/cell_gen_fragment.h (limited to 'src/gallium/drivers/cell/spu/spu_per_fragment_op.h') diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h index c0ca201e1d..a62530c64d 100644 --- a/src/gallium/drivers/cell/common.h +++ b/src/gallium/drivers/cell/common.h @@ -97,6 +97,7 @@ #define CELL_CMD_STATE_LOGICOP 21 #define CELL_CMD_VS_EXECUTE 22 #define CELL_CMD_FLUSH_BUFFER_RANGE 23 +#define CELL_CMD_STATE_FRAGMENT_OPS 24 #define CELL_NUM_BUFFERS 4 @@ -112,30 +113,43 @@ /** */ -struct cell_command_depth_stencil_alpha_test { +struct cell_command_depth_stencil_alpha_test +{ uint64_t base; /**< Effective address of code start. */ unsigned size; /**< Size in bytes of SPE code. */ unsigned read_depth; /**< Flag: should depth be read? */ unsigned read_stencil; /**< Flag: should stencil be read? */ + struct pipe_depth_stencil_alpha_state state; }; /** * Upload code to perform framebuffer blend operation */ -struct cell_command_blend { +struct cell_command_blend +{ uint64_t base; /**< Effective address of code start. */ unsigned size; /**< Size in bytes of SPE code. */ unsigned read_fb; /**< Flag: should framebuffer be read? */ }; -struct cell_command_logicop { +struct cell_command_logicop +{ uint64_t base; /**< Effective address of code start. */ unsigned size; /**< Size in bytes of SPE code. */ }; +#define SPU_MAX_FRAGMENT_OPS_INSTS 64 + +struct cell_command_fragment_ops +{ + uint64_t opcode; /**< CELL_CMD_STATE_FRAGMENT_OPS */ + unsigned code[SPU_MAX_FRAGMENT_OPS_INSTS]; +}; + + /** * Tell SPUs about the framebuffer size, location */ diff --git a/src/gallium/drivers/cell/ppu/Makefile b/src/gallium/drivers/cell/ppu/Makefile index 25473e200c..b5a6fcb8de 100644 --- a/src/gallium/drivers/cell/ppu/Makefile +++ b/src/gallium/drivers/cell/ppu/Makefile @@ -25,6 +25,7 @@ SOURCES = \ cell_context.c \ cell_draw_arrays.c \ cell_flush.c \ + cell_gen_fragment.c \ cell_state_derived.c \ cell_state_emit.c \ cell_state_per_fragment.c \ diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c new file mode 100644 index 0000000000..df29476be6 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -0,0 +1,530 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + + +/** + * Generate SPU per-fragment code (actually per-quad code). + * \author Brian Paul + */ + + +#include "pipe/p_defines.h" +#include "pipe/p_state.h" +#include "rtasm/rtasm_ppc_spe.h" +#include "cell_context.h" +#include "cell_gen_fragment.h" + + + +/** Do extra optimizations? */ +#define OPTIMIZATIONS 1 + + +/** + * Generate SPE code to perform Z/depth testing. + * + * \param dsa Gallium depth/stencil/alpha state to gen code for + * \param f SPE function to append instruction onto. + * \param mask_reg register containing quad/pixel "alive" mask (in/out) + * \param ifragZ_reg register containing integer fragment Z values (in) + * \param ifbZ_reg register containing integer frame buffer Z values (in/out) + * \param zmask_reg register containing result of Z test/comparison (out) + */ +static void +gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, + struct spe_function *f, + int mask_reg, int ifragZ_reg, int ifbZ_reg, int zmask_reg) +{ + ASSERT(dsa->depth.enabled); + + switch (dsa->depth.func) { + case PIPE_FUNC_EQUAL: + /* zmask = (ifragZ == ref) */ + spe_ceq(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_NOTEQUAL: + /* zmask = (ifragZ == ref) */ + spe_ceq(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_GREATER: + /* zmask = (ifragZ > ref) */ + spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_LESS: + /* zmask = (ref > ifragZ) */ + spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_LEQUAL: + /* zmask = (ifragZ > ref) */ + spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_GEQUAL: + /* zmask = (ref > ifragZ) */ + spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_NEVER: + spe_il(f, mask_reg, 0); /* mask = {0,0,0,0} */ + spe_move(f, zmask_reg, mask_reg); /* zmask = mask */ + break; + + case PIPE_FUNC_ALWAYS: + /* mask unchanged */ + spe_il(f, zmask_reg, ~0); /* zmask = {~0,~0,~0,~0} */ + break; + + default: + ASSERT(0); + break; + } + + if (dsa->depth.writemask) { + /* + * If (ztest passed) { + * framebufferZ = fragmentZ; + * } + * OR, + * framebufferZ = (ztest_passed ? fragmentZ : framebufferZ; + */ + spe_selb(f, ifbZ_reg, ifbZ_reg, ifragZ_reg, mask_reg); + } +} + + +/** + * Generate SPE code to perform alpha testing. + * + * \param dsa Gallium depth/stencil/alpha state to gen code for + * \param f SPE function to append instruction onto. + * \param mask_reg register containing quad/pixel "alive" mask (in/out) + * \param fragA_reg register containing four fragment alpha values (in) + */ +static void +gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, + struct spe_function *f, int mask_reg, int fragA_reg) +{ + int ref_reg = spe_allocate_available_register(f); + int amask_reg = spe_allocate_available_register(f); + + ASSERT(dsa->alpha.enabled); + + if ((dsa->alpha.func != PIPE_FUNC_NEVER) && + (dsa->alpha.func != PIPE_FUNC_ALWAYS)) { + /* load/splat the alpha reference float value */ + spe_load_float(f, ref_reg, dsa->alpha.ref); + } + + /* emit code to do the alpha comparison, updating 'mask' */ + switch (dsa->alpha.func) { + case PIPE_FUNC_EQUAL: + /* amask = (fragA == ref) */ + spe_fceq(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_NOTEQUAL: + /* amask = (fragA == ref) */ + spe_fceq(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_GREATER: + /* amask = (fragA > ref) */ + spe_fcgt(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_LESS: + /* amask = (ref > fragA) */ + spe_fcgt(f, amask_reg, ref_reg, fragA_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_LEQUAL: + /* amask = (fragA > ref) */ + spe_fcgt(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_GEQUAL: + /* amask = (ref > fragA) */ + spe_fcgt(f, amask_reg, ref_reg, fragA_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_NEVER: + spe_il(f, mask_reg, 0); /* mask = [0,0,0,0] */ + break; + + case PIPE_FUNC_ALWAYS: + /* no-op, mask unchanged */ + break; + + default: + ASSERT(0); + break; + } + +#if OPTIMIZATIONS + /* if mask == {0,0,0,0} we're all done, return */ + { + /* re-use amask reg here */ + int tmp_reg = amask_reg; + /* tmp[0] = (mask[0] | mask[1] | mask[2] | mask[3]) */ + spe_orx(f, tmp_reg, mask_reg); + /* if tmp[0] == 0 then return from function call */ + spe_biz(f, tmp_reg, SPE_REG_RA, 0, 0); + } +#endif + + spe_release_register(f, ref_reg); + spe_release_register(f, amask_reg); +} + + + +/** + * Generate SPE code to implement the fragment operations (alpha test, + * depth test, stencil test, blending, colormask, and final + * framebuffer write) as specified by the current context state. + * + * Logically, this code will be called after running the fragment + * shader. But under some circumstances we could run some of this + * code before the fragment shader to cull fragments/quads that are + * totally occluded/discarded. + * + * XXX we only support PIPE_FORMAT_Z24S8_UNORM z/stencil buffer right now. + * + * See the spu_default_fragment_ops() function to see how the per-fragment + * operations would be done with ordinary C code. + * The code we generate here though has no branches, is SIMD, etc and + * should be much faster. + * + * \param cell the rendering context (in) + * \param f the generated function (out) + */ +void +gen_fragment_function(struct cell_context *cell, struct spe_function *f) +{ + const struct pipe_depth_stencil_alpha_state *dsa = + &cell->depth_stencil->base; + const struct pipe_blend_state *blend = &cell->blend->base; + + /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */ + const int x_reg = 3; /* uint */ + const int y_reg = 4; /* uint */ + const int color_tile_reg = 5; /* tile_t * */ + const int depth_tile_reg = 6; /* tile_t * */ + const int fragZ_reg = 7; /* vector float */ + const int fragR_reg = 8; /* vector float */ + const int fragG_reg = 9; /* vector float */ + const int fragB_reg = 10; /* vector float */ + const int fragA_reg = 11; /* vector float */ + const int mask_reg = 12; /* vector uint */ + + /* offset of quad from start of tile + * XXX assuming 4-byte pixels for color AND Z/stencil!!!! + */ + int quad_offset_reg; + + int fbRGBA_reg; /**< framebuffer's RGBA colors for quad */ + int fbZS_reg; /**< framebuffer's combined z/stencil values for quad */ + + spe_init_func(f, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + spe_allocate_register(f, x_reg); + spe_allocate_register(f, y_reg); + spe_allocate_register(f, color_tile_reg); + spe_allocate_register(f, depth_tile_reg); + spe_allocate_register(f, fragZ_reg); + spe_allocate_register(f, fragR_reg); + spe_allocate_register(f, fragG_reg); + spe_allocate_register(f, fragB_reg); + spe_allocate_register(f, fragA_reg); + spe_allocate_register(f, mask_reg); + + quad_offset_reg = spe_allocate_available_register(f); + fbRGBA_reg = spe_allocate_available_register(f); + fbZS_reg = spe_allocate_available_register(f); + + /* compute offset of quad from start of tile, in bytes */ + { + int x2_reg = spe_allocate_available_register(f); + int y2_reg = spe_allocate_available_register(f); + + ASSERT(TILE_SIZE == 32); + + spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */ + spe_rotmi(f, y2_reg, y_reg, -1); /* y2 = y / 2 */ + spe_shli(f, y2_reg, y2_reg, 4); /* y2 *= 16 */ + spe_a(f, quad_offset_reg, y2_reg, x2_reg); /* offset = y2 + x2 */ + spe_shli(f, quad_offset_reg, quad_offset_reg, 4); /* offset *= 16 */ + + spe_release_register(f, x2_reg); + spe_release_register(f, y2_reg); + } + + + if (dsa->alpha.enabled) { + gen_alpha_test(dsa, f, mask_reg, fragA_reg); + } + + if (dsa->depth.enabled || dsa->stencil[0].enabled) { + const enum pipe_format zs_format = cell->framebuffer.zsbuf->format; + boolean write_depth_stencil; + + int fbZ_reg = spe_allocate_available_register(f); /* Z values */ + int fbS_reg = spe_allocate_available_register(f); /* Stencil values */ + + /* fetch quad of depth/stencil values from tile at (x,y) */ + /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + + if (dsa->depth.enabled) { + /* Extract Z bits from fbZS_reg into fbZ_reg */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + int mask_reg = spe_allocate_available_register(f); + spe_fsmbi(f, mask_reg, 0x7777); /* mask[0,1,2,3] = 0x00ffffff */ + spe_and(f, fbZ_reg, fbZS_reg, mask_reg); /* fbZ = fbZS & mask */ + spe_release_register(f, mask_reg); + /* OK, fbZ_reg has four 24-bit Z values now */ + } + else { + /* XXX handle other z/stencil formats */ + ASSERT(0); + } + + /* Convert fragZ values from float[4] to uint[4] */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM || + zs_format == PIPE_FORMAT_Z24S8_UNORM || + zs_format == PIPE_FORMAT_Z24X8_UNORM) { + /* 24-bit Z values */ + int scale_reg = spe_allocate_available_register(f); + + /* scale_reg[0,1,2,3] = float(2^24-1) */ + spe_load_float(f, scale_reg, (float) 0xffffff); + + /* XXX these two instructions might be combined */ + spe_fm(f, fragZ_reg, fragZ_reg, scale_reg); /* fragZ *= scale */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 0); /* fragZ = (int) fragZ */ + + spe_release_register(f, scale_reg); + } + else { + /* XXX handle 16-bit Z format */ + ASSERT(0); + } + } + + if (dsa->stencil[0].enabled) { + /* Extract Stencil bit sfrom fbZS_reg into fbS_reg */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + /* XXX extract with a shift */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || + zs_format == PIPE_FORMAT_Z24X8_UNORM) { + /* XXX extract with a mask */ + ASSERT(0); + } + } + + + if (dsa->stencil[0].enabled) { + /* XXX this may involve depth testing too */ + // gen_stencil_test(dsa, f, ... ); + ASSERT(0); + } + else if (dsa->depth.enabled) { + int zmask_reg = spe_allocate_available_register(f); + gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); + spe_release_register(f, zmask_reg); + } + + /* do we need to write Z and/or Stencil back into framebuffer? */ + write_depth_stencil = (dsa->depth.writemask | + dsa->stencil[0].write_mask | + dsa->stencil[1].write_mask); + + if (write_depth_stencil) { + /* Merge latest Z and Stencil values into fbZS_reg. + * fbZ_reg has four Z vals in bits [23..0] or bits [15..0]. + * fbS_reg has four 8-bit Z values in bits [7..0]. + */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } + else if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_Z16_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_S8_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else { + /* bad zs_format */ + ASSERT(0); + } + + /* Store: memory[depth_tile_reg + quad_offset_reg] = fbZS */ + spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + } + + spe_release_register(f, fbZ_reg); + spe_release_register(f, fbS_reg); + } + + + /* Get framebuffer quad/colors. We'll need these for blending, + * color masking, and to obey the quad/pixel mask. + * Load: fbRGBA_reg = memory[color_tile + quad_offset] + * Note: if mask={~0,~0,~0,~0} and we're not blending or colormasking + * we could skip this load. + */ + spe_lqx(f, fbRGBA_reg, color_tile_reg, quad_offset_reg); + + + if (blend->blend_enable) { + /* convert packed tile colors in fbRGBA_reg to float[4] vectors */ + + // gen_blend_code(blend, f, mask_reg, ... ); + + } + + + + /* + * Write fragment colors to framebuffer/tile. + * This involves converting the fragment colors from float[4] to the + * tile's specific format and obeying the quad/pixel mask. + */ + { + const enum pipe_format color_format = cell->framebuffer.cbufs[0]->format; + int rgba_reg = spe_allocate_available_register(f); + + /* Convert float[4] in [0.0,1.0] to int[4] in [0,~0], with clamping */ + spe_cfltu(f, fragR_reg, fragR_reg, 32); + spe_cfltu(f, fragG_reg, fragG_reg, 32); + spe_cfltu(f, fragB_reg, fragB_reg, 32); + spe_cfltu(f, fragA_reg, fragA_reg, 32); + + /* Shift most the significant bytes to least the significant positions. + * I.e.: reg = reg >> 24 + */ + spe_rotmi(f, fragR_reg, fragR_reg, -24); + spe_rotmi(f, fragG_reg, fragG_reg, -24); + spe_rotmi(f, fragB_reg, fragB_reg, -24); + spe_rotmi(f, fragA_reg, fragA_reg, -24); + + /* Shift the color bytes according to the surface format */ + if (color_format == PIPE_FORMAT_A8R8G8B8_UNORM) { + spe_roti(f, fragG_reg, fragG_reg, 8); /* green <<= 8 */ + spe_roti(f, fragR_reg, fragR_reg, 16); /* red <<= 16 */ + spe_roti(f, fragA_reg, fragA_reg, 24); /* alpha <<= 24 */ + } + else if (color_format == PIPE_FORMAT_B8G8R8A8_UNORM) { + spe_roti(f, fragR_reg, fragR_reg, 8); /* red <<= 8 */ + spe_roti(f, fragG_reg, fragG_reg, 16); /* green <<= 16 */ + spe_roti(f, fragB_reg, fragB_reg, 24); /* blue <<= 24 */ + } + else { + ASSERT(0); + } + + /* Merge red, green, blue, alpha registers to make packed RGBA colors. + * Eg: after shifting according to color_format we might have: + * R = {0x00ff0000, 0x00110000, 0x00220000, 0x00330000} + * G = {0x0000ff00, 0x00004400, 0x00005500, 0x00006600} + * B = {0x000000ff, 0x00000077, 0x00000088, 0x00000099} + * A = {0xff000000, 0xaa000000, 0xbb000000, 0xcc000000} + * OR-ing all those together gives us four packed colors: + * RGBA = {0xffffffff, 0xaa114477, 0xbb225588, 0xcc336699} + */ + spe_or(f, rgba_reg, fragR_reg, fragG_reg); + spe_or(f, rgba_reg, rgba_reg, fragB_reg); + spe_or(f, rgba_reg, rgba_reg, fragA_reg); + + /* Mix fragment colors with framebuffer colors using the quad/pixel mask: + * if (mask[i]) + * rgba[i] = rgba[i]; + * else + * rgba[i] = framebuffer[i]; + */ + spe_selb(f, rgba_reg, fbRGBA_reg, rgba_reg, mask_reg); + + /* Store updated quad in tile: + * memory[color_tile + quad_offset] = rgba_reg; + */ + spe_stqx(f, rgba_reg, color_tile_reg, quad_offset_reg); + + spe_release_register(f, rgba_reg); + } + + printf("gen_fragment_ops nr instructions: %u\n", f->num_inst); + + spe_bi(f, SPE_REG_RA, 0, 0); /* return from function call */ + + + spe_release_register(f, fbRGBA_reg); + spe_release_register(f, fbZS_reg); + spe_release_register(f, quad_offset_reg); +} + diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h new file mode 100644 index 0000000000..0ea0fc690c --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h @@ -0,0 +1,38 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef CELL_GEN_FRAGMENT_H +#define CELL_GEN_FRAGMENT_H + + +extern void +gen_fragment_function(struct cell_context *cell, struct spe_function *f); + + +#endif /* CELL_GEN_FRAGMENT_H */ + diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c index f2feaa329a..06777aac14 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -27,6 +27,7 @@ #include "util/u_memory.h" #include "cell_context.h" +#include "cell_gen_fragment.h" #include "cell_state.h" #include "cell_state_emit.h" #include "cell_state_per_fragment.h" @@ -83,6 +84,29 @@ cell_emit_state(struct cell_context *cell) fb->depth_format = zbuf ? zbuf->format : PIPE_FORMAT_NONE; fb->width = cell->framebuffer.width; fb->height = cell->framebuffer.height; +#if 0 + printf("EMIT color format %s\n", pf_name(fb->color_format)); + printf("EMIT depth format %s\n", pf_name(fb->depth_format)); +#endif + } + + + if (cell->dirty & (CELL_NEW_FRAMEBUFFER | CELL_NEW_DEPTH_STENCIL)) { + /* XXX we don't want to always do codegen here. We should have + * a hash/lookup table to cache previous results... + */ + struct cell_command_fragment_ops *fops + = cell_batch_alloc(cell, sizeof(*fops)); + struct spe_function spe_code; + + /* generate new code */ + gen_fragment_function(cell, &spe_code); + /* put the new code into the batch buffer */ + fops->opcode = CELL_CMD_STATE_FRAGMENT_OPS; + memcpy(&fops->code, spe_code.store, + SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + /* free codegen buffer */ + spe_release_func(&spe_code); } if (cell->dirty & CELL_NEW_BLEND) { @@ -90,8 +114,7 @@ cell_emit_state(struct cell_context *cell) if (cell->blend != NULL) { blend.base = (intptr_t) cell->blend->code.store; - blend.size = (char *) cell->blend->code.csr - - (char *) cell->blend->code.store; + blend.size = cell->blend->code.num_inst * SPE_INST_SIZE; blend.read_fb = TRUE; } else { @@ -108,10 +131,10 @@ cell_emit_state(struct cell_context *cell) if (cell->depth_stencil != NULL) { dsat.base = (intptr_t) cell->depth_stencil->code.store; - dsat.size = (char *) cell->depth_stencil->code.csr - - (char *) cell->depth_stencil->code.store; + dsat.size = cell->depth_stencil->code.num_inst * SPE_INST_SIZE; dsat.read_depth = TRUE; dsat.read_stencil = FALSE; + dsat.state = cell->depth_stencil->base; } else { dsat.base = 0; diff --git a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c index 705867107b..78cb446c14 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c @@ -1158,7 +1158,7 @@ cell_generate_alpha_blend(struct cell_blend_state *cb) static int PC_OFFSET(const struct spe_function *f, const void *d) { - const intptr_t pc = (intptr_t) f->csr; + const intptr_t pc = (intptr_t) &f->store[f->num_inst]; const intptr_t ea = ~0x0f & (intptr_t) d; return (ea - pc) >> 2; diff --git a/src/gallium/drivers/cell/spu/Makefile b/src/gallium/drivers/cell/spu/Makefile index d49abb2e82..e285ae9fdb 100644 --- a/src/gallium/drivers/cell/spu/Makefile +++ b/src/gallium/drivers/cell/spu/Makefile @@ -43,7 +43,7 @@ INCLUDE_DIRS = \ $(SPU_CC) $(SPU_CFLAGS) -c $< .c.s: - $(SPU_CC) $(SPU_CFLAGS) -S $< + $(SPU_CC) $(SPU_CFLAGS) -O3 -S $< # The .a file will be linked into the main/PPU executable diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c index c4236817a9..4e0ec15925 100644 --- a/src/gallium/drivers/cell/spu/spu_main.c +++ b/src/gallium/drivers/cell/spu/spu_main.c @@ -34,6 +34,7 @@ #include "spu_main.h" #include "spu_render.h" +#include "spu_per_fragment_op.h" #include "spu_texture.h" #include "spu_tile.h" //#include "spu_test.h" @@ -46,7 +47,7 @@ /* helpful headers: /usr/lib/gcc/spu/4.1.1/include/spu_mfcio.h -/opt/ibm/cell-sdk/prototype/sysroot/usr/include/libmisc.h +/opt/cell/sdk/usr/include/libmisc.h */ boolean Debug = FALSE; @@ -226,6 +227,24 @@ cmd_release_verts(const struct cell_command_release_verts *release) } +/** + * Process a CELL_CMD_STATE_FRAGMENT_OPS command. + * This involves installing new fragment ops SPU code. + * If this function is never called, we'll use a regular C fallback function + * for fragment processing. + */ +static void +cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) +{ + if (Debug) + printf("SPU %u: CMD_STATE_FRAGMENT_OPS\n", spu.init.id); + /* Copy SPU code from batch buffer to spu buffer */ + memcpy(spu.fragment_ops.code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4); + /* Point function pointer at new code */ + spu.fragment_ops.func = (spu_fragment_ops_func) spu.fragment_ops.code; +} + + static void cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) { @@ -257,6 +276,8 @@ cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) break; case PIPE_FORMAT_Z24S8_UNORM: case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_Z24X8_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: spu.fb.zsize = 4; spu.fb.zscale = (float) 0x00ffffffu; break; @@ -282,6 +303,8 @@ cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) } +#define NEW_FRAGMENT_FUNCTION 01 + static void cmd_state_blend(const struct cell_command_blend *state) { @@ -302,7 +325,9 @@ cmd_state_blend(const struct cell_command_blend *state) wait_on_mask(1 << TAG_BATCH_BUFFER); spu.blend = (blend_func) fb_blend_code_buffer; spu.read_fb = state->read_fb; - } else { + } + else + { spu.read_fb = FALSE; } } @@ -326,7 +351,9 @@ cmd_state_depth_stencil(const struct cell_command_depth_stencil_alpha_test *stat 0, /* tid */ 0 /* rid */); wait_on_mask(1 << TAG_BATCH_BUFFER); - } else { + } + else + { /* If there is no code, emit a return instruction. */ depth_stencil_code_buffer[0] = 0x35; @@ -338,12 +365,14 @@ cmd_state_depth_stencil(const struct cell_command_depth_stencil_alpha_test *stat spu.frag_test = (frag_test_func) depth_stencil_code_buffer; spu.read_depth = state->read_depth; spu.read_stencil = state->read_stencil; + spu.depth_stencil_alpha = state->state; } static void cmd_state_logicop(const struct cell_command_logicop * code) { +#if !NEW_FRAGMENT_FUNCTION mfc_get(logicop_code_buffer, (unsigned int) code->base, /* src */ code->size, @@ -353,6 +382,7 @@ cmd_state_logicop(const struct cell_command_logicop * code) wait_on_mask(1 << TAG_BATCH_BUFFER); spu.logicop = (logicop_func) logicop_code_buffer; +#endif } @@ -455,7 +485,9 @@ cmd_finish(void) /** - * Execute a batch of commands + * Execute a batch of commands which was sent to us by the PPU. + * See the cell_emit_state.c code to see where the commands come from. + * * The opcode param encodes the location of the buffer and its size. */ static void @@ -519,6 +551,14 @@ cmd_batch(uint opcode) pos += pos_incr; } break; + case CELL_CMD_STATE_FRAGMENT_OPS: + { + struct cell_command_fragment_ops *fops + = (struct cell_command_fragment_ops *) &buffer[pos]; + cmd_state_fragment_ops(fops); + pos += sizeof(*fops) / 8; + } + break; case CELL_CMD_RELEASE_VERTS: { struct cell_command_release_verts *release @@ -680,6 +720,11 @@ one_time_init(void) memset(spu.ctile_status, TILE_STATUS_DEFINED, sizeof(spu.ctile_status)); memset(spu.ztile_status, TILE_STATUS_DEFINED, sizeof(spu.ztile_status)); invalidate_tex_cache(); + + /* Install default/fallback fragment processing function. + * This will normally be overriden by a code-gen'd function. + */ + spu.fragment_ops.func = spu_fallback_fragment_ops; } diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h index c2a53c9dcf..7ab34f5222 100644 --- a/src/gallium/drivers/cell/spu/spu_main.h +++ b/src/gallium/drivers/cell/spu/spu_main.h @@ -91,6 +91,24 @@ typedef struct spu_blend_results (*logicop_func)( typedef vector float (*sample_texture_func)(uint unit, vector float texcoord); + +typedef void (*spu_fragment_ops_func)(uint x, uint y, + tile_t *colorTile, + tile_t *depthStencilTile, + vector float fragZ, + vector float fragRed, + vector float fragGreen, + vector float fragBlue, + vector float fragAlpha, + vector unsigned int mask); + +struct spu_fragment_ops +{ + uint code[SPU_MAX_FRAGMENT_OPS_INSTS]; + spu_fragment_ops_func func; /**< Current fragment ops function */ +} ALIGN16_ATTRIB; + + struct spu_framebuffer { void *color_start; /**< addr of color surface in main memory */ void *depth_start; /**< addr of depth surface in main memory */ @@ -127,6 +145,9 @@ struct spu_global struct cell_init_info init; struct spu_framebuffer fb; + + struct pipe_depth_stencil_alpha_state depth_stencil_alpha; + boolean read_depth; boolean read_stencil; frag_test_func frag_test; /**< Current depth/stencil test code */ @@ -142,6 +163,8 @@ struct spu_global struct vertex_info vertex_info; + struct spu_fragment_ops fragment_ops; + /* XXX more state to come */ diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c index 29dc07a2e8..ffc596aa62 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c @@ -29,8 +29,11 @@ * \author Ian Romanick */ + +#include #include "pipe/p_format.h" #include "spu_main.h" +#include "spu_colorpack.h" #include "spu_per_fragment_op.h" #define ZERO 0x80 @@ -90,7 +93,8 @@ read_ds_quad(tile_t *tile, unsigned x, unsigned y, break; } - case PIPE_FORMAT_S8Z24_UNORM: { + case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: { qword *ptr = (qword *) &tile->ui4[iy][ix]; *depth = si_and(*ptr, si_fsmbi(0x7777)); @@ -153,7 +157,8 @@ write_ds_quad(tile_t *buffer, unsigned x, unsigned y, break; } - case PIPE_FORMAT_S8Z24_UNORM: { + case PIPE_FORMAT_S8Z24_UNORM: + case PIPE_FORMAT_X8Z24_UNORM: { qword *ptr = (qword *) &buffer->ui4[iy][ix]; /* form select mask = 0111,0111,0111,0111 */ qword mask = si_fsmbi(0x7777); @@ -217,3 +222,225 @@ spu_do_depth_stencil(int x, int y, return result.mask; } + + + + +/** + * Called by rasterizer for each quad after the shader has run. This + * is a fallback/debug function. In reality we'll use a generated + * function produced by the PPU. But this function is useful for + * debug/validation. + */ +void +spu_fallback_fragment_ops(uint x, uint y, + tile_t *colorTile, + tile_t *depthStencilTile, + vector float fragZ, + vector float fragRed, + vector float fragGreen, + vector float fragBlue, + vector float fragAlpha, + vector unsigned int mask) +{ + vector float frag_soa[4], frag_aos[4]; + unsigned int c0, c1, c2, c3; + + /* do alpha test */ + if (spu.depth_stencil_alpha.alpha.enabled) { + vector float ref = spu_splats(spu.depth_stencil_alpha.alpha.ref); + vector unsigned int amask; + + switch (spu.depth_stencil_alpha.alpha.func) { + case PIPE_FUNC_LESS: + amask = spu_cmpgt(ref, fragAlpha); /* mask = (fragAlpha < ref) */ + break; + case PIPE_FUNC_GREATER: + amask = spu_cmpgt(fragAlpha, ref); /* mask = (fragAlpha > ref) */ + break; + case PIPE_FUNC_GEQUAL: + amask = spu_cmpgt(ref, fragAlpha); + amask = spu_nor(amask, amask); + break; + case PIPE_FUNC_LEQUAL: + amask = spu_cmpgt(fragAlpha, ref); + amask = spu_nor(amask, amask); + break; + case PIPE_FUNC_EQUAL: + amask = spu_cmpeq(ref, fragAlpha); + break; + case PIPE_FUNC_NOTEQUAL: + amask = spu_cmpeq(ref, fragAlpha); + amask = spu_nor(amask, amask); + break; + case PIPE_FUNC_ALWAYS: + amask = spu_splats(0xffffffffU); + break; + case PIPE_FUNC_NEVER: + amask = spu_splats( 0x0U); + break; + default: + ; + } + + mask = spu_and(mask, amask); + } + + /* Z and/or stencil testing... */ + if (spu.depth_stencil_alpha.depth.enabled || + spu.depth_stencil_alpha.stencil[0].enabled) { + + /* get four Z/Stencil values from tile */ + vector unsigned int mask24 = spu_splats((unsigned int)0x00ffffffU); + vector unsigned int ifbZS = depthStencilTile->ui4[y/2][x/2]; + vector unsigned int ifbZ = spu_and(ifbZS, mask24); + vector unsigned int ifbS = spu_andc(ifbZS, mask24); + + if (spu.depth_stencil_alpha.stencil[0].enabled) { + /* do stencil test */ + ASSERT(spu.fb.depth_format == PIPE_FORMAT_S8Z24_UNORM); + + } + else if (spu.depth_stencil_alpha.depth.enabled) { + /* do depth test */ + + ASSERT(spu.fb.depth_format == PIPE_FORMAT_S8Z24_UNORM || + spu.fb.depth_format == PIPE_FORMAT_X8Z24_UNORM); + + vector unsigned int ifragZ; + vector unsigned int zmask; + + /* convert four fragZ from float to uint */ + fragZ = spu_mul(fragZ, spu_splats((float) 0xffffff)); + ifragZ = spu_convtu(fragZ, 0); + + /* do depth comparison, setting zmask with results */ + switch (spu.depth_stencil_alpha.depth.func) { + case PIPE_FUNC_LESS: + zmask = spu_cmpgt(ifbZ, ifragZ); /* mask = (ifragZ < ifbZ) */ + break; + case PIPE_FUNC_GREATER: + zmask = spu_cmpgt(ifragZ, ifbZ); /* mask = (ifbZ > ifragZ) */ + break; + case PIPE_FUNC_GEQUAL: + zmask = spu_cmpgt(ifbZ, ifragZ); + zmask = spu_nor(zmask, zmask); + break; + case PIPE_FUNC_LEQUAL: + zmask = spu_cmpgt(ifragZ, ifbZ); + zmask = spu_nor(zmask, zmask); + break; + case PIPE_FUNC_EQUAL: + zmask = spu_cmpeq(ifbZ, ifragZ); + break; + case PIPE_FUNC_NOTEQUAL: + zmask = spu_cmpeq(ifbZ, ifragZ); + zmask = spu_nor(zmask, zmask); + break; + case PIPE_FUNC_ALWAYS: + zmask = spu_splats(0xffffffffU); + break; + case PIPE_FUNC_NEVER: + zmask = spu_splats( 0x0U); + break; + default: + ; + } + + mask = spu_and(mask, zmask); + + /* merge framebuffer Z and fragment Z according to the mask */ + ifbZ = spu_or(spu_and(ifragZ, mask), + spu_andc(ifbZ, mask)); + } + + if (spu_extract(spu_orx(mask), 0)) { + /* put new fragment Z/Stencil values back into Z/Stencil tile */ + depthStencilTile->ui4[y/2][x/2] = spu_or(ifbZ, ifbS); + + spu.cur_ztile_status = TILE_STATUS_DIRTY; + } + } + + /* XXX do blending here */ + + /* XXX do colormask test here */ + + + if (spu_extract(spu_orx(mask), 0)) { + spu.cur_ctile_status = TILE_STATUS_DIRTY; + } + else { + return; + } + + /* convert RRRR,GGGG,BBBB,AAAA to RGBA,RGBA,RGBA,RGBA */ +#if 0 + { + vector float frag_soa[4]; + frag_soa[0] = fragRed; + frag_soa[1] = fragGreen; + frag_soa[2] = fragBlue; + frag_soa[3] = fragAlpha; + _transpose_matrix4x4(frag_aos, frag_soa); + } +#else + /* short-cut relying on function parameter layout: */ + _transpose_matrix4x4(frag_aos, &fragRed); + (void) fragGreen; + (void) fragBlue; +#endif + + switch (spu.fb.color_format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + c0 = spu_pack_A8R8G8B8(frag_aos[0]); + c1 = spu_pack_A8R8G8B8(frag_aos[1]); + c2 = spu_pack_A8R8G8B8(frag_aos[2]); + c3 = spu_pack_A8R8G8B8(frag_aos[3]); + break; + + case PIPE_FORMAT_B8G8R8A8_UNORM: + c0 = spu_pack_B8G8R8A8(frag_aos[0]); + c1 = spu_pack_B8G8R8A8(frag_aos[1]); + c2 = spu_pack_B8G8R8A8(frag_aos[2]); + c3 = spu_pack_B8G8R8A8(frag_aos[3]); + break; + default: + fprintf(stderr, "SPU: Bad pixel format in spu_default_fragment_ops\n"); + ASSERT(0); + } + +#if 0 + /* + * Quad layout: + * +--+--+ + * |p0|p1| + * +--+--+ + * |p2|p3| + * +--+--+ + */ + if (spu_extract(mask, 0)) + colorTile->ui[y+0][x+0] = c0; + if (spu_extract(mask, 1)) + colorTile->ui[y+0][x+1] = c1; + if (spu_extract(mask, 2)) + colorTile->ui[y+1][x+0] = c2; + if (spu_extract(mask, 3)) + colorTile->ui[y+1][x+1] = c3; +#else + /* + * Quad layout: + * +--+--+--+--+ + * |p0|p1|p2|p3| + * +--+--+--+--+ + */ + if (spu_extract(mask, 0)) + colorTile->ui[y][x*2] = c0; + if (spu_extract(mask, 1)) + colorTile->ui[y][x*2+1] = c1; + if (spu_extract(mask, 2)) + colorTile->ui[y][x*2+2] = c2; + if (spu_extract(mask, 3)) + colorTile->ui[y][x*2+3] = c3; +#endif +} diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h index 6571258699..ffadf0661c 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h @@ -29,4 +29,15 @@ extern qword spu_do_depth_stencil(int x, int y, qword frag_mask, qword frag_depth, qword frag_alpha, qword facing); +extern void +spu_fallback_fragment_ops(uint x, uint y, + tile_t *colorTile, + tile_t *depthStencilTile, + vector float fragZ, + vector float fragRed, + vector float fragGreen, + vector float fragBlue, + vector float fragAlpha, + vector unsigned int mask); + #endif /* SPU_PER_FRAGMENT_OP */ diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index a3ea0a3e69..71ef6ca24f 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -297,9 +297,12 @@ emit_quad( int x, int y, mask_t mask ) sp->quad.first->run(sp->quad.first, &setup.quad); #else +#define NEW_FRAGMENT_FUNCTION 01 +#if !NEW_FRAGMENT_FUNCTION if (spu.read_depth) { mask = do_depth_test(x, y, mask); } +#endif /* If any bits in mask are set... */ if (spu_extract(spu_orx(mask), 0)) { @@ -308,6 +311,7 @@ emit_quad( int x, int y, mask_t mask ) vector float colors[4]; spu.cur_ctile_status = TILE_STATUS_DIRTY; + spu.cur_ztile_status = TILE_STATUS_DIRTY; if (spu.texture[0].start) { /* texture mapping */ @@ -355,6 +359,29 @@ emit_quad( int x, int y, mask_t mask ) } +#if NEW_FRAGMENT_FUNCTION + { + /* Convert fragment data from AoS to SoA format. + * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) + * This is temporary! + */ + vector float soa_frag[4]; + _transpose_matrix4x4(soa_frag, colors); + + float4 fragZ; + + fragZ.v = eval_z((float) x, (float) y); + + /* Do all per-fragment/quad operations here, including: + * alpha test, z test, stencil test, blend and framebuffer writing. + */ + spu.fragment_ops.func(ix, iy, &spu.ctile, &spu.ztile, + fragZ.v, + soa_frag[0], soa_frag[1], + soa_frag[2], soa_frag[3], + mask); + } +#else /* Convert fragment data from AoS to SoA format. * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) */ @@ -405,6 +432,9 @@ emit_quad( int x, int y, mask_t mask ) spu.ctile.ui[iy+0][ix+1] = spu_extract((vec_uint4) result.g, 0); spu.ctile.ui[iy+1][ix+0] = spu_extract((vec_uint4) result.b, 0); spu.ctile.ui[iy+1][ix+1] = spu_extract((vec_uint4) result.a, 0); + +#endif /* NEW_FRAGMENT_FUNCTION */ + } #endif } diff --git a/src/gallium/winsys/xlib/xm_api.c b/src/gallium/winsys/xlib/xm_api.c index b010513107..28bd6ceab4 100644 --- a/src/gallium/winsys/xlib/xm_api.c +++ b/src/gallium/winsys/xlib/xm_api.c @@ -349,12 +349,17 @@ create_xmesa_buffer(XMesaDrawable d, BufferType type, if (vis->mesa_visual.depthBits == 0) depthFormat = PIPE_FORMAT_NONE; +#ifdef GALLIUM_CELL /* XXX temporary for Cell! */ + else + depthFormat = PIPE_FORMAT_S8Z24_UNORM; +#else else if (vis->mesa_visual.depthBits <= 16) - depthFormat = PIPE_FORMAT_Z16_UNORM; + depthFormat = PIPE_FORMAT_Z16UNORM; else if (vis->mesa_visual.depthBits <= 24) depthFormat = PIPE_FORMAT_S8Z24_UNORM; else depthFormat = PIPE_FORMAT_Z32_UNORM; +#endif if (vis->mesa_visual.stencilBits == 8) { if (depthFormat == PIPE_FORMAT_S8Z24_UNORM) diff --git a/src/gallium/winsys/xlib/xm_winsys.c b/src/gallium/winsys/xlib/xm_winsys.c index 5e9a1f92f1..c4a30d3702 100644 --- a/src/gallium/winsys/xlib/xm_winsys.c +++ b/src/gallium/winsys/xlib/xm_winsys.c @@ -275,6 +275,39 @@ xm_buffer_destroy(struct pipe_winsys *pws, } +/** + * For Cell. Basically, rearrange the pixels/quads from this layout: + * +--+--+--+--+ + * |p0|p1|p2|p3|.... + * +--+--+--+--+ + * + * to this layout: + * +--+--+ + * |p0|p1|.... + * +--+--+ + * |p2|p3| + * +--+--+ + */ +static void +twiddle_tile(uint *tile) +{ + uint tile2[TILE_SIZE * TILE_SIZE]; + int y, x; + + for (y = 0; y < TILE_SIZE; y+=2) { + for (x = 0; x < TILE_SIZE; x+=2) { + int k = 4 * (y/2 * TILE_SIZE/2 + x/2); + tile2[y * TILE_SIZE + (x + 0)] = tile[k]; + tile2[y * TILE_SIZE + (x + 1)] = tile[k+1]; + tile2[(y + 1) * TILE_SIZE + (x + 0)] = tile[k+2]; + tile2[(y + 1) * TILE_SIZE + (x + 1)] = tile[k+3]; + } + } + memcpy(tile, tile2, sizeof(tile2)); +} + + + /** * Display a surface that's in a tiled configuration. That is, all the * pixels for a TILE_SIZExTILE_SIZE block are contiguous in memory. @@ -321,6 +354,8 @@ xmesa_display_surface_tiled(XMesaBuffer b, const struct pipe_surface *surf) ximage->data = (char *) xm_buf->data + offset; + twiddle_tile((uint *) ximage->data); + if (XSHM_ENABLED(xm_buf)) { #if defined(USE_XSHM) && !defined(XFree86Server) XShmPutImage(b->xm_visual->display, b->drawable, b->gc, -- cgit v1.2.3 From 701fcee65db6b72f98e926d838956bbcc54f1cc6 Mon Sep 17 00:00:00 2001 From: Brian Paul Date: Wed, 10 Sep 2008 18:51:43 -0600 Subject: cell: remove old per-fragment code, replace with all new code --- src/gallium/drivers/cell/spu/spu_per_fragment_op.c | 236 +++------------------ src/gallium/drivers/cell/spu/spu_per_fragment_op.h | 47 ++-- src/gallium/drivers/cell/spu/spu_tri.c | 96 --------- 3 files changed, 48 insertions(+), 331 deletions(-) (limited to 'src/gallium/drivers/cell/spu/spu_per_fragment_op.h') diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c index ffc596aa62..9ed5fc50cd 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c @@ -1,32 +1,32 @@ -/* - * (C) Copyright IBM Corporation 2008 +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the "Software"), - * to deal in the Software without restriction, including without limitation - * on the rights to use, copy, modify, merge, publish, distribute, sub - * license, and/or sell copies of the Software, and to permit persons to whom - * the Software is furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice (including the next - * paragraph) shall be included in all copies or substantial portions of the - * Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL - * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, - * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR - * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE - * USE OR OTHER DEALINGS IN THE SOFTWARE. - */ + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ /** - * \file spu_per_fragment_op.c - * SPU implementation various per-fragment operations. - * - * \author Ian Romanick + * \author Brian Paul */ @@ -36,194 +36,6 @@ #include "spu_colorpack.h" #include "spu_per_fragment_op.h" -#define ZERO 0x80 - - -/** - * Get a "quad" of four fragment Z/stencil values from the given tile. - * \param tile the tile of Z/stencil values - * \param x, y location of the quad in the tile, in pixels - * \param depth_format format of the tile's data - * \param detph returns four depth values - * \param stencil returns four stencil values - */ -static void -read_ds_quad(tile_t *tile, unsigned x, unsigned y, - enum pipe_format depth_format, qword *depth, - qword *stencil) -{ - const int ix = x / 2; - const int iy = y / 2; - - switch (depth_format) { - case PIPE_FORMAT_Z16_UNORM: { - qword *ptr = (qword *) &tile->us8[iy][ix / 2]; - - const qword shuf_vec = (qword) { - ZERO, ZERO, 0, 1, ZERO, ZERO, 2, 3, - ZERO, ZERO, 4, 5, ZERO, ZERO, 6, 7 - }; - - /* At even X values we want the first 4 shorts, and at odd X values we - * want the second 4 shorts. - */ - qword bias = (qword) spu_splats((unsigned char) ((ix & 0x01) << 3)); - qword bias_mask = si_fsmbi(0x3333); - qword sv = si_a(shuf_vec, si_and(bias_mask, bias)); - - *depth = si_shufb(*ptr, *ptr, sv); - *stencil = si_il(0); - break; - } - - case PIPE_FORMAT_Z32_UNORM: { - qword *ptr = (qword *) &tile->ui4[iy][ix]; - - *depth = *ptr; - *stencil = si_il(0); - break; - } - - case PIPE_FORMAT_Z24S8_UNORM: { - qword *ptr = (qword *) &tile->ui4[iy][ix]; - qword mask = si_fsmbi(0xEEEE); - - *depth = si_rotmai(si_and(*ptr, mask), -8); - *stencil = si_andc(*ptr, mask); - break; - } - - case PIPE_FORMAT_S8Z24_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: { - qword *ptr = (qword *) &tile->ui4[iy][ix]; - - *depth = si_and(*ptr, si_fsmbi(0x7777)); - *stencil = si_andi(si_roti(*ptr, 8), 0x0ff); - break; - } - - default: - ASSERT(0); - break; - } -} - - -/** - * Put a quad of Z/stencil values into a tile. - * \param tile the tile of Z/stencil values to write into - * \param x, y location of the quad in the tile, in pixels - * \param depth_format format of the tile's data - * \param detph depth values to store - * \param stencil stencil values to store - */ -static void -write_ds_quad(tile_t *buffer, unsigned x, unsigned y, - enum pipe_format depth_format, - qword depth, qword stencil) -{ - const int ix = x / 2; - const int iy = y / 2; - - (void) stencil; - - switch (depth_format) { - case PIPE_FORMAT_Z16_UNORM: { - qword *ptr = (qword *) &buffer->us8[iy][ix / 2]; - - qword sv = ((ix & 0x01) == 0) - ? (qword) { 2, 3, 6, 7, 10, 11, 14, 15, - 24, 25, 26, 27, 28, 29, 30, 31 } - : (qword) { 16, 17, 18, 19, 20 , 21, 22, 23, - 2, 3, 6, 7, 10, 11, 14, 15 }; - *ptr = si_shufb(depth, *ptr, sv); - break; - } - - case PIPE_FORMAT_Z32_UNORM: { - qword *ptr = (qword *) &buffer->ui4[iy][ix]; - *ptr = depth; - break; - } - - case PIPE_FORMAT_Z24S8_UNORM: { - qword *ptr = (qword *) &buffer->ui4[iy][ix]; - /* form select mask = 1110,1110,1110,1110 */ - qword mask = si_fsmbi(0xEEEE); - /* depth[i] = depth[i] << 8 */ - depth = si_shli(depth, 8); - /* *ptr[i] = depth[i][31:8] | stencil[i][7:0] */ - *ptr = si_selb(stencil, depth, mask); - break; - } - - case PIPE_FORMAT_S8Z24_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: { - qword *ptr = (qword *) &buffer->ui4[iy][ix]; - /* form select mask = 0111,0111,0111,0111 */ - qword mask = si_fsmbi(0x7777); - /* stencil[i] = stencil[i] << 24 */ - stencil = si_shli(stencil, 24); - /* *ptr[i] = stencil[i][31:24] | depth[i][23:0] */ - *ptr = si_selb(stencil, depth, mask); - break; - } - - default: - ASSERT(0); - break; - } -} - - -/** - * Do depth/stencil/alpha test for a "quad" of 4 fragments. - * \param x,y location of quad within tile - * \param frag_mask indicates which fragments are "alive" - * \param frag_depth four fragment depth values - * \param frag_alpha four fragment alpha values - * \param facing front/back facing for four fragments (1=front, 0=back) - */ -qword -spu_do_depth_stencil(int x, int y, - qword frag_mask, qword frag_depth, qword frag_alpha, - qword facing) -{ - struct spu_frag_test_results result; - qword pixel_depth; - qword pixel_stencil; - - /* All of this preable code (everthing before the call to frag_test) should - * be generated on the PPU and upload to the SPU. - */ - if (spu.read_depth || spu.read_stencil) { - read_ds_quad(&spu.ztile, x, y, spu.fb.depth_format, - &pixel_depth, &pixel_stencil); - } - - /* convert floating point Z values to 32-bit uint */ - - /* frag_depth *= spu.fb.zscale */ - frag_depth = si_fm(frag_depth, (qword)spu_splats(spu.fb.zscale)); - /* frag_depth = uint(frag_depth) */ - frag_depth = si_cfltu(frag_depth, 0); - - result = (*spu.frag_test)(frag_mask, pixel_depth, pixel_stencil, - frag_depth, frag_alpha, facing); - - - /* This code (everthing after the call to frag_test) should - * be generated on the PPU and upload to the SPU. - */ - if (spu.read_depth || spu.read_stencil) { - write_ds_quad(&spu.ztile, x, y, spu.fb.depth_format, - result.depth, result.stencil); - } - - return result.mask; -} - - /** diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h index ffadf0661c..f817abf046 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h @@ -1,33 +1,33 @@ -/* - * (C) Copyright IBM Corporation 2008 +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. * All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a - * copy of this software and associated documentation files (the "Software"), - * to deal in the Software without restriction, including without limitation - * on the rights to use, copy, modify, merge, publish, distribute, sub - * license, and/or sell copies of the Software, and to permit persons to whom - * the Software is furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice (including the next - * paragraph) shall be included in all copies or substantial portions of the - * Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL - * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, - * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR - * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE - * USE OR OTHER DEALINGS IN THE SOFTWARE. - */ + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ #ifndef SPU_PER_FRAGMENT_OP #define SPU_PER_FRAGMENT_OP -extern qword -spu_do_depth_stencil(int x, int y, qword frag_mask, qword frag_depth, - qword frag_alpha, qword facing); extern void spu_fallback_fragment_ops(uint x, uint y, @@ -40,4 +40,5 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fragAlpha, vector unsigned int mask); + #endif /* SPU_PER_FRAGMENT_OP */ diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index 71ef6ca24f..a5bf3270c7 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -38,7 +38,6 @@ #include "spu_texture.h" #include "spu_tile.h" #include "spu_tri.h" -#include "spu_per_fragment_op.h" /** Masks are uint[4] vectors with each element being 0 or 0xffffffff */ @@ -255,31 +254,6 @@ eval_z(float x, float y) } -static INLINE mask_t -do_depth_test(int x, int y, mask_t quadmask) -{ - float4 zvals; - mask_t mask; - - if (spu.fb.depth_format == PIPE_FORMAT_NONE) - return quadmask; - - zvals.v = eval_z((float) x, (float) y); - - mask = (mask_t) spu_do_depth_stencil(x - setup.cliprect_minx, - y - setup.cliprect_miny, - (qword) quadmask, - (qword) zvals.v, - (qword) spu_splats((unsigned char) 0x0ffu), - (qword) spu_splats((unsigned int) 0x01u)); - - if (spu_extract(spu_orx(mask), 0)) - spu.cur_ztile_status = TILE_STATUS_DIRTY; - - return mask; -} - - /** * Emit a quad (pass to next stage). No clipping is done. * Note: about 1/5 to 1/7 of the time, mask is zero and this function @@ -289,21 +263,6 @@ do_depth_test(int x, int y, mask_t quadmask) static INLINE void emit_quad( int x, int y, mask_t mask ) { -#if 0 - struct softpipe_context *sp = setup.softpipe; - setup.quad.x0 = x; - setup.quad.y0 = y; - setup.quad.mask = mask; - sp->quad.first->run(sp->quad.first, &setup.quad); -#else - -#define NEW_FRAGMENT_FUNCTION 01 -#if !NEW_FRAGMENT_FUNCTION - if (spu.read_depth) { - mask = do_depth_test(x, y, mask); - } -#endif - /* If any bits in mask are set... */ if (spu_extract(spu_orx(mask), 0)) { const int ix = x - setup.cliprect_minx; @@ -359,7 +318,6 @@ emit_quad( int x, int y, mask_t mask ) } -#if NEW_FRAGMENT_FUNCTION { /* Convert fragment data from AoS to SoA format. * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) @@ -381,62 +339,8 @@ emit_quad( int x, int y, mask_t mask ) soa_frag[2], soa_frag[3], mask); } -#else - /* Convert fragment data from AoS to SoA format. - * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) - */ - qword soa_frag[4]; - _transpose_matrix4x4((vec_float4 *) soa_frag, colors); - - /* Read the current framebuffer values. - */ - const qword pix[4] = { - (qword) spu_splats(spu.ctile.ui[iy+0][ix+0]), - (qword) spu_splats(spu.ctile.ui[iy+0][ix+1]), - (qword) spu_splats(spu.ctile.ui[iy+1][ix+0]), - (qword) spu_splats(spu.ctile.ui[iy+1][ix+1]), - }; - - qword soa_pix[4]; - - if (spu.read_fb) { - /* Convert pixel data from AoS to SoA format. - * I.e. (RGBA,RGBA,RGBA,RGBA) -> (RRRR,GGGG,BBBB,AAAA) - */ - vec_float4 aos_pix[4] = { - spu_unpack_A8R8G8B8(spu.ctile.ui[iy+0][ix+0]), - spu_unpack_A8R8G8B8(spu.ctile.ui[iy+0][ix+1]), - spu_unpack_A8R8G8B8(spu.ctile.ui[iy+1][ix+0]), - spu_unpack_A8R8G8B8(spu.ctile.ui[iy+1][ix+1]), - }; - - _transpose_matrix4x4((vec_float4 *) soa_pix, aos_pix); - } - - - struct spu_blend_results result = - (*spu.blend)(soa_frag[0], soa_frag[1], soa_frag[2], soa_frag[3], - soa_pix[0], soa_pix[1], soa_pix[2], soa_pix[3], - spu.const_blend_color[0], spu.const_blend_color[1], - spu.const_blend_color[2], spu.const_blend_color[3]); - - - /* Convert final pixel data from SoA to AoS format. - * I.e. (RRRR,GGGG,BBBB,AAAA) -> (RGBA,RGBA,RGBA,RGBA) - */ - result = (*spu.logicop)(pix[0], pix[1], pix[2], pix[3], - result.r, result.g, result.b, result.a, - (qword) mask); - - spu.ctile.ui[iy+0][ix+0] = spu_extract((vec_uint4) result.r, 0); - spu.ctile.ui[iy+0][ix+1] = spu_extract((vec_uint4) result.g, 0); - spu.ctile.ui[iy+1][ix+0] = spu_extract((vec_uint4) result.b, 0); - spu.ctile.ui[iy+1][ix+1] = spu_extract((vec_uint4) result.a, 0); - -#endif /* NEW_FRAGMENT_FUNCTION */ } -#endif } -- cgit v1.2.3 From afaa53040bd01ca86762e7d7b1a5a65810767921 Mon Sep 17 00:00:00 2001 From: Robert Ellison Date: Fri, 3 Oct 2008 18:00:43 -0600 Subject: CELL: changes to generate SPU code for stenciling This set of code changes are for stencil code generation support. Both one-sided and two-sided stenciling are supported. In addition to the raw code generation changes, these changes had to be made elsewhere in the system: - Added new "register set" feature to the SPE assembly generation. A "register set" is a way to allocate multiple registers and free them all at the same time, delegating register allocation management to the spe_function unit. It's quite useful in complex register allocation schemes (like stenciling). - Added and improved SPE macro calculations. These are operations between registers and unsigned integer immediates. In many cases, the calculation can be performed with a single instruction; the macros will generate the single instruction if possible, or generate a register load and register-to-register operation if not. These macro functions are: spe_load_uint() (which has new ways to load a value in a single instruction), spe_and_uint(), spe_xor_uint(), spe_compare_equal_uint(), and spe_compare_greater_uint(). - Added facing to fragment generation. While rendering, the rasterizer needs to be able to determine front- and back-facing fragments, in order to correctly apply two-sided stencil. That requires these changes: - Added front_winding field to the cell_command_render block, so that the state tracker could communicate to the rasterizer what it considered to be the front-facing direction. - Added fragment facing as an input to the fragment function. - Calculated facing is passed during emit_quad(). --- src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c | 246 +++++- src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h | 41 +- src/gallium/drivers/cell/common.h | 1 + src/gallium/drivers/cell/ppu/cell_gen_fragment.c | 881 ++++++++++++++++++--- src/gallium/drivers/cell/ppu/cell_render.c | 1 + src/gallium/drivers/cell/ppu/cell_vbuf.c | 1 + src/gallium/drivers/cell/spu/spu_main.h | 3 +- src/gallium/drivers/cell/spu/spu_per_fragment_op.c | 19 +- src/gallium/drivers/cell/spu/spu_per_fragment_op.h | 3 +- src/gallium/drivers/cell/spu/spu_render.c | 4 +- src/gallium/drivers/cell/spu/spu_tri.c | 35 +- src/gallium/drivers/cell/spu/spu_tri.h | 2 +- 12 files changed, 1091 insertions(+), 146 deletions(-) (limited to 'src/gallium/drivers/cell/spu/spu_per_fragment_op.h') diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c index 491141f190..8a87e9abb1 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c @@ -359,14 +359,21 @@ void _name (struct spe_function *p, int imm) \ */ void spe_init_func(struct spe_function *p, unsigned code_size) { + register unsigned int i; + p->store = align_malloc(code_size, 16); p->num_inst = 0; p->max_inst = code_size / SPE_INST_SIZE; + p->set_count = 0; + memset(p->regs, 0, SPE_NUM_REGS * sizeof(p->regs[0])); + /* Conservatively treat R0 - R2 and R80 - R127 as non-volatile. */ - p->regs[0] = ~7; - p->regs[1] = (1U << (80 - 64)) - 1; + p->regs[0] = p->regs[1] = p->regs[2] = 1; + for (i = 80; i <= 127; i++) { + p->regs[i] = 1; + } p->print = false; p->indent = 0; @@ -398,12 +405,8 @@ int spe_allocate_available_register(struct spe_function *p) { unsigned i; for (i = 0; i < SPE_NUM_REGS; i++) { - const uint64_t mask = (1ULL << (i % 64)); - const unsigned idx = i / 64; - - assert(idx < 2); - if ((p->regs[idx] & mask) != 0) { - p->regs[idx] &= ~mask; + if (p->regs[i] == 0) { + p->regs[i] = 1; return i; } } @@ -417,31 +420,68 @@ int spe_allocate_available_register(struct spe_function *p) */ int spe_allocate_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) != 0); - - p->regs[idx] &= ~(1ULL << bit); + assert(p->regs[reg] == 0); + p->regs[reg] = 1; return reg; } /** - * Mark the given SPE register as "unallocated". + * Mark the given SPE register as "unallocated". Note that this should + * only be used on registers allocated in the current register set; an + * assertion will fail if an attempt is made to deallocate a register + * allocated in an earlier register set. */ void spe_release_register(struct spe_function *p, int reg) { - const unsigned idx = reg / 64; - const unsigned bit = reg % 64; + assert(reg < SPE_NUM_REGS); + assert(p->regs[reg] == 1); - assert(idx < 2); + p->regs[reg] = 0; +} - assert(reg < SPE_NUM_REGS); - assert((p->regs[idx] & (1ULL << bit)) == 0); +/** + * Start a new set of registers. This can be called if + * it will be difficult later to determine exactly what + * registers were actually allocated during a code generation + * sequence, and you really just want to deallocate all of them. + */ +void spe_allocate_register_set(struct spe_function *p) +{ + register unsigned int i; + + /* Keep track of the set count. If it ever wraps around to 0, + * we're in trouble. + */ + p->set_count++; + assert(p->set_count > 0); + + /* Increment the allocation count of all registers currently + * allocated. Then any registers that are allocated in this set + * will be the only ones with a count of 1; they'll all be released + * when the register set is released. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) p->regs[i]++; + } +} + +void spe_release_register_set(struct spe_function *p) +{ + unsigned int i; + + /* If the set count drops below zero, we're in trouble. */ + assert(p->set_count > 0); + p->set_count--; - p->regs[idx] |= (1ULL << bit); + /* Drop the allocation level of all registers. Any allocated + * during this register set will drop to 0 and then become + * available. + */ + for (i = 0; i < SPE_NUM_REGS; i++) { + if (p->regs[i] > 0) p->regs[i]--; + } } @@ -603,8 +643,10 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui) { /* If the whole value is in the lower 18 bits, use ila, which * doesn't sign-extend. Otherwise, if the two halfwords of - * the constant are identical, use ilh. Otherwise, we have - * to use ilhu followed by iohl. + * the constant are identical, use ilh. Otherwise, if every byte of + * the desired value is 0x00 or 0xff, we can use Form Select Mask for + * Bytes Immediate (fsmbi) to load the value in a single instruction. + * Otherwise, in the general case, we have to use ilhu followed by iohl. */ if ((ui & 0xfffc0000) == ui) { spe_ila(p, rT, ui); @@ -612,13 +654,171 @@ void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui) else if ((ui >> 16) == (ui & 0xffff)) { spe_ilh(p, rT, ui & 0xffff); } + else if ( + ((ui & 0x000000ff) == 0 || (ui & 0x000000ff) == 0x000000ff) && + ((ui & 0x0000ff00) == 0 || (ui & 0x0000ff00) == 0x0000ff00) && + ((ui & 0x00ff0000) == 0 || (ui & 0x00ff0000) == 0x00ff0000) && + ((ui & 0xff000000) == 0 || (ui & 0xff000000) == 0xff000000) + ) { + unsigned int mask = 0; + /* fsmbi duplicates each bit in the given mask eight times, + * using a 16-bit value to initialize a 16-byte quadword. + * Each 4-bit nybble of the mask corresponds to a full word + * of the result; look at the value and figure out the mask + * (replicated for each word in the quadword), and then + * form the "select mask" to get the value. + */ + if ((ui & 0x000000ff) == 0x000000ff) mask |= 0x1111; + if ((ui & 0x0000ff00) == 0x0000ff00) mask |= 0x2222; + if ((ui & 0x00ff0000) == 0x00ff0000) mask |= 0x4444; + if ((ui & 0xff000000) == 0xff000000) mask |= 0x8888; + spe_fsmbi(p, rT, mask); + } else { + /* The general case: this usually uses two instructions, but + * may use only one if the low-order 16 bits of each word are 0. + */ spe_ilhu(p, rT, ui >> 16); if (ui & 0xffff) spe_iohl(p, rT, ui & 0xffff); } } +/* This function is constructed identically to spe_sor_uint() below. + * Changes to one should be made in the other. + */ +void spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either And Byte Immediate + * (which uses the same constant across each byte), And Halfword Immediate + * (which sign-extends a 10-bit immediate to 16 bits and uses that + * across each halfword), or And Word Immediate (which sign-extends + * a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + register unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use And Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_andi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use And Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_andhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the And Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_andbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_and(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + +/* This function is constructed identically to spe_and_uint() above. + * Changes to one should be made in the other. + */ +void spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If we can, emit a single instruction, either Exclusive Or Byte + * Immediate (which uses the same constant across each byte), Exclusive + * Or Halfword Immediate (which sign-extends a 10-bit immediate to + * 16 bits and uses that across each halfword), or Exclusive Or Word + * Immediate (which sign-extends a 10-bit immediate to 32 bits). + * + * Otherwise, we'll need to use a temporary register. + */ + register unsigned int tmp; + + /* If the upper 23 bits are all 0s or all 1s, sign extension + * will work and we can use Exclusive Or Word Immediate + */ + tmp = ui & 0xfffffe00; + if (tmp == 0xfffffe00 || tmp == 0) { + spe_xori(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric along halfword boundaries and + * the upper 7 bits of each halfword are all 0s or 1s, we + * can use Exclusive Or Halfword Immediate + */ + tmp = ui & 0xfe00fe00; + if ((tmp == 0xfe00fe00 || tmp == 0) && ((ui >> 16) == (ui & 0x0000ffff))) { + spe_xorhi(p, rT, rA, ui & 0x000003ff); + return; + } + + /* If the ui field is symmetric in each byte, then we can use + * the Exclusive Or Byte Immediate instruction. + */ + tmp = ui & 0x000000ff; + if ((ui >> 24) == tmp && ((ui >> 16) & 0xff) == tmp && ((ui >> 8) & 0xff) == tmp) { + spe_xorbi(p, rT, rA, tmp); + return; + } + + /* Otherwise, we'll have to use a temporary register. */ + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_xor(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); +} + +void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 9 bits or less, it fits inside a + * Compare Equal Word Immediate instruction. + */ + if ((ui & 0x000001ff) == ui) { + spe_ceqi(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_ceq(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); + } +} + +void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui) +{ + /* If the comparison value is 10 bits or less, it fits inside a + * Compare Logical Greater Than Word Immediate instruction. + */ + if ((ui & 0x000003ff) == ui) { + spe_clgti(p, rT, rA, ui); + } + /* Otherwise, we're going to have to load a word first. */ + else { + unsigned int tmp_reg = spe_allocate_available_register(p); + spe_load_uint(p, tmp_reg, ui); + spe_clgt(p, rT, rA, tmp_reg); + spe_release_register(p, tmp_reg); + } +} void spe_splat(struct spe_function *p, unsigned rT, unsigned rA) diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h index 61c7edeb60..cd2e245409 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h @@ -53,17 +53,26 @@ struct spe_function uint num_inst; uint max_inst; - /** - * Mask of used / unused registers - * - * Each set bit corresponds to an available register. Each cleared bit - * corresponds to an allocated register. + /** + * The "set count" reflects the number of nested register sets + * are allowed. In the unlikely case that we exceed the set count, + * register allocation will start to be confused, which is critical + * enough that we check for it. + */ + unsigned char set_count; + + /** + * Flags for used and unused registers. Each byte corresponds to a + * register; a 0 in that byte means that the register is available. + * A value of 1 means that the register was allocated in the current + * register set. Any other value N means that the register was allocated + * N register sets ago. * * \sa * spe_allocate_register, spe_allocate_available_register, - * spe_release_register + * spe_allocate_register_set, spe_release_register_set, spe_release_register, */ - uint64_t regs[SPE_NUM_REGS / 64]; + unsigned char regs[SPE_NUM_REGS]; boolean print; /**< print/dump instructions as they're emitted? */ int indent; /**< number of spaces to indent */ @@ -77,6 +86,8 @@ extern unsigned spe_code_size(const struct spe_function *p); extern int spe_allocate_available_register(struct spe_function *p); extern int spe_allocate_register(struct spe_function *p, int reg); extern void spe_release_register(struct spe_function *p, int reg); +extern void spe_allocate_register_set(struct spe_function *p); +extern void spe_release_register_set(struct spe_function *p); extern void spe_print_code(struct spe_function *p, boolean enable); extern void spe_indent(struct spe_function *p, int spaces); @@ -307,6 +318,22 @@ spe_load_int(struct spe_function *p, unsigned rT, int i); extern void spe_load_uint(struct spe_function *p, unsigned rT, unsigned int ui); +/** And immediate value into rT. */ +extern void +spe_and_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Xor immediate value into rT. */ +extern void +spe_xor_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare equal with immediate value. */ +extern void +spe_compare_equal_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + +/** Compare greater with immediate value. */ +extern void +spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsigned int ui); + /** Replicate word 0 of rA across rT. */ extern void spe_splat(struct spe_function *p, unsigned rT, unsigned rA); diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h index 99329fd8e2..c223bc1744 100644 --- a/src/gallium/drivers/cell/common.h +++ b/src/gallium/drivers/cell/common.h @@ -227,6 +227,7 @@ struct cell_command_render float xmin, ymin, xmax, ymax; /* XXX another dummy field */ uint min_index; boolean inline_verts; + uint front_winding; /* the rasterizer needs to be able to determine facing to apply front/back-facing stencil */ }; diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c index 653afc235d..f920ae13b4 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -54,10 +54,12 @@ * \param ifragZ_reg register containing integer fragment Z values (in) * \param ifbZ_reg register containing integer frame buffer Z values (in/out) * \param zmask_reg register containing result of Z test/comparison (out) + * + * Returns true if the Z-buffer needs to be updated. */ -static void -gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, - struct spe_function *f, +static boolean +gen_depth_test(struct spe_function *f, + const struct pipe_depth_stencil_alpha_state *dsa, int mask_reg, int ifragZ_reg, int ifbZ_reg, int zmask_reg) { /* NOTE: we use clgt below, not cgt, because we want to compare _unsigned_ @@ -132,7 +134,10 @@ gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, * framebufferZ = (ztest_passed ? fragmentZ : framebufferZ; */ spe_selb(f, ifbZ_reg, ifbZ_reg, ifragZ_reg, mask_reg); + return true; } + + return false; } @@ -238,22 +243,34 @@ gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, * it and have to allocate and load it again unnecessarily. */ static inline void -setup_const_register(struct spe_function *f, boolean *is_already_set, unsigned int *r, float value) +setup_optional_register(struct spe_function *f, boolean *is_already_set, unsigned int *r) { if (*is_already_set) return; *r = spe_allocate_available_register(f); - spe_load_float(f, *r, value); - *is_already_set = true; } static inline void -release_const_register(struct spe_function *f, boolean *is_already_set, unsigned int r) +release_optional_register(struct spe_function *f, boolean *is_already_set, unsigned int r) { if (!*is_already_set) return; spe_release_register(f, r); *is_already_set = false; } +static inline void +setup_const_register(struct spe_function *f, boolean *is_already_set, unsigned int *r, float value) +{ + if (*is_already_set) return; + setup_optional_register(f, is_already_set, r); + spe_load_float(f, *r, value); +} + +static inline void +release_const_register(struct spe_function *f, boolean *is_already_set, unsigned int r) +{ + release_optional_register(f, is_already_set, r); +} + /** * Generate SPE code to implement the given blend mode for a quad of pixels. * \param f SPE function to append instruction onto. @@ -1117,6 +1134,633 @@ gen_colormask(struct spe_function *f, spe_release_register(f, colormask_reg); } +/* This function is annoyingly similar to gen_depth_test(), above, except + * that instead of comparing two varying values (i.e. fragment and buffer), + * we're comparing a varying value with a static value. As such, we have + * access to the Compare Immediate instructions where we don't in + * gen_depth_test(), which is what makes us very different. + * + * The return value in the stencil_pass_reg is a bitmask of valid + * fragments that also passed the stencil test. The bitmask of valid + * fragments that failed would be found in (mask_reg & ~stencil_pass_reg). + */ +static void +gen_stencil_test(struct spe_function *f, const struct pipe_stencil_state *state, + unsigned int mask_reg, unsigned int fbS_reg, + unsigned int stencil_pass_reg) +{ + /* Generate code that puts the set of passing fragments into the stencil_pass_reg + * register, taking into account whether each fragment was active to begin with. + */ + switch (state->func) { + case PIPE_FUNC_EQUAL: + /* stencil_pass = mask & (s == reference) */ + spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + /* stencil_fail = mask & ~stencil_pass */ + break; + + case PIPE_FUNC_NOTEQUAL: + /* stencil_pass = mask & ~(s == reference) */ + spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + break; + + case PIPE_FUNC_GREATER: + /* stencil_pass = mask & (s > reference) */ + spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + break; + + case PIPE_FUNC_LESS: { + /* stencil_pass = mask & (reference > s) */ + /* There's no convenient Compare Less Than Immediate instruction, so + * we'll have to do this one the harder way, by loading a register and + * comparing directly. Compare Logical Greater Than Word (clgt) + * treats its operands as unsigned - no sign extension. + */ + unsigned int tmp_reg = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value); + spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg); + spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + break; + } + + case PIPE_FUNC_LEQUAL: + /* stencil_pass = mask & (s <= reference) = mask & ~(s > reference) */ + spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, state->ref_value); + spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + break; + + case PIPE_FUNC_GEQUAL: { + /* stencil_pass = mask & (s >= reference) = mask & ~(reference > s) */ + /* As above, we have to do this by loading a register */ + unsigned int tmp_reg = spe_allocate_available_register(f); + spe_load_uint(f, tmp_reg, state->ref_value); + spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg); + spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg); + spe_release_register(f, tmp_reg); + break; + } + + case PIPE_FUNC_NEVER: + /* stencil_pass = mask & 0 = 0 */ + spe_load_uint(f, stencil_pass_reg, 0); + spe_move(f, stencil_pass_reg, mask_reg); /* zmask = mask */ + break; + + case PIPE_FUNC_ALWAYS: + /* stencil_pass = mask & 1 = mask */ + spe_move(f, stencil_pass_reg, mask_reg); + break; + } + + /* The fragments that passed the stencil test are now in stencil_pass_reg. + * The fragments that failed would be (mask_reg & ~stencil_pass_reg). + */ +} + +/* This function generates code that calculates a set of new stencil values + * given the earlier values and the operation to apply. It does not + * apply any tests. It is intended to be called up to 3 times + * (for the stencil fail operation, for the stencil pass-z fail operation, + * and for the stencil pass-z pass operation) to collect up to three + * possible sets of values, and for the caller to combine them based + * on the result of the tests. + * + * stencil_max_value should be (2^n - 1) where n is the number of bits + * in the stencil buffer - in other words, it should be usable as a mask. + */ +static void +gen_stencil_values(struct spe_function *f, unsigned int stencil_op, + unsigned int stencil_ref_value, unsigned int stencil_max_value, + unsigned int fbS_reg, unsigned int newS_reg) +{ + /* The code below assumes that newS_reg and fbS_reg are not the same + * register; if they can be, the calculations below will have to use + * an additional temporary register. For now, mark the assumption + * with an assertion that will fail if they are the same. + */ + ASSERT(fbS_reg != newS_reg); + + /* The code also assumes the the stencil_max_value is of the form + * 2^n-1 and can therefore be used as a mask for the valid bits in + * addition to a maximum. Make sure this is the case as well. + * The clever math below exploits the fact that incrementing a + * binary number serves to flip all the bits of a number starting at + * the LSB and continuing to (and including) the first zero bit + * found. That means that a number and its increment will always + * have at least one bit in common (the high order bit, if nothing + * else) *unless* the number is zero, *or* the number is of a form + * consisting of some number of 1s in the low-order bits followed + * by nothing but 0s in the high-order bits. The latter case + * implies it's of the form 2^n-1. + */ + ASSERT(stencil_max_value > 0 && ((stencil_max_value + 1) & stencil_max_value) == 0); + + switch(stencil_op) { + case PIPE_STENCIL_OP_KEEP: + /* newS = S */ + spe_move(f, newS_reg, fbS_reg); + break; + + case PIPE_STENCIL_OP_ZERO: + /* newS = 0 */ + spe_zero(f, newS_reg); + break; + + case PIPE_STENCIL_OP_REPLACE: + /* newS = stencil reference value */ + spe_load_uint(f, newS_reg, stencil_ref_value); + break; + + case PIPE_STENCIL_OP_INCR: { + /* newS = (s == max ? max : s + 1) */ + unsigned int equals_reg = spe_allocate_available_register(f); + + spe_compare_equal_uint(f, equals_reg, fbS_reg, stencil_max_value); + /* Add Word Immediate computes rT = rA + 10-bit signed immediate */ + spe_ai(f, newS_reg, fbS_reg, 1); + /* Select from the current value or the new value based on the equality test */ + spe_selb(f, newS_reg, fbS_reg, newS_reg, equals_reg); + + spe_release_register(f, equals_reg); + break; + } + case PIPE_STENCIL_OP_DECR: { + /* newS = (s == 0 ? 0 : s - 1) */ + unsigned int equals_reg = spe_allocate_available_register(f); + + spe_compare_equal_uint(f, equals_reg, fbS_reg, 0); + /* Add Word Immediate with a (-1) value works */ + spe_ai(f, newS_reg, fbS_reg, -1); + /* Select from the current value or the new value based on the equality test */ + spe_selb(f, newS_reg, fbS_reg, newS_reg, equals_reg); + + spe_release_register(f, equals_reg); + break; + } + case PIPE_STENCIL_OP_INCR_WRAP: + /* newS = (s == max ? 0 : s + 1), but since max is 2^n-1, we can + * do a normal add and mask off the correct bits + */ + spe_ai(f, newS_reg, fbS_reg, 1); + spe_and_uint(f, newS_reg, newS_reg, stencil_max_value); + break; + + case PIPE_STENCIL_OP_DECR_WRAP: + /* newS = (s == 0 ? max : s - 1), but we'll pull the same mask trick as above */ + spe_ai(f, newS_reg, fbS_reg, -1); + spe_and_uint(f, newS_reg, newS_reg, stencil_max_value); + break; + + case PIPE_STENCIL_OP_INVERT: + /* newS = ~s. We take advantage of the mask/max value to invert only + * the valid bits for the field so we don't have to do an extra "and". + */ + spe_xor_uint(f, newS_reg, fbS_reg, stencil_max_value); + break; + + default: + ASSERT(0); + } +} + + +/* This function generates code to get all the necessary possible + * stencil values. For each of the output registers (fail_reg, + * zfail_reg, and zpass_reg), it either allocates a new register + * and calculates a new set of values based on the stencil operation, + * or it reuses a register allocation and calculation done for an + * earlier (matching) operation, or it reuses the fbS_reg register + * (if the stencil operation is KEEP, which doesn't change the + * stencil buffer). + * + * Since this function allocates a variable number of registers, + * to avoid incurring complex logic to free them, they should + * be allocated after a spe_allocate_register_set() call + * and released by the corresponding spe_release_register_set() call. + */ +static void +gen_get_stencil_values(struct spe_function *f, const struct pipe_depth_stencil_alpha_state *dsa, + unsigned int fbS_reg, + unsigned int *fail_reg, unsigned int *zfail_reg, + unsigned int *zpass_reg, unsigned int *back_fail_reg, + unsigned int *back_zfail_reg, unsigned int *back_zpass_reg) +{ + unsigned zfail_op, back_zfail_op; + + /* Stenciling had better be enabled here */ + ASSERT(dsa->stencil[0].enabled); + + /* If the depth test is not enabled, it is treated as though it always + * passes. In particular, that means that the "zfail_op" (and the backfacing + * counterpart, if active) are not considered - a failing stencil test will + * trigger the "fail_op", and a passing stencil test will trigger the + * "zpass_op". + * + * By overriding the operations in this case to be PIPE_STENCIL_OP_KEEP, + * we keep them from being calculated. + */ + if (dsa->depth.enabled) { + zfail_op = dsa->stencil[0].zfail_op; + back_zfail_op = dsa->stencil[1].zfail_op; + } + else { + zfail_op = PIPE_STENCIL_OP_KEEP; + back_zfail_op = PIPE_STENCIL_OP_KEEP; + } + + /* One-sided or front-facing stencil */ + if (dsa->stencil[0].fail_op == PIPE_STENCIL_OP_KEEP) { + *fail_reg = fbS_reg; + } + else { + *fail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[0].fail_op, dsa->stencil[0].ref_value, + 0xff, fbS_reg, *fail_reg); + } + + if (zfail_op == PIPE_STENCIL_OP_KEEP) { + *zfail_reg = fbS_reg; + } + else if (zfail_op == dsa->stencil[0].fail_op) { + *zfail_reg = *fail_reg; + } + else { + *zfail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[0].zfail_op, dsa->stencil[0].ref_value, + 0xff, fbS_reg, *zfail_reg); + } + + if (dsa->stencil[0].zpass_op == PIPE_STENCIL_OP_KEEP) { + *zpass_reg = fbS_reg; + } + else if (dsa->stencil[0].zpass_op == dsa->stencil[0].fail_op) { + *zpass_reg = *fail_reg; + } + else if (dsa->stencil[0].zpass_op == zfail_op) { + *zpass_reg = *zfail_reg; + } + else { + *zpass_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[0].zpass_op, dsa->stencil[0].ref_value, + 0xff, fbS_reg, *zpass_reg); + } + + /* If two-sided stencil is enabled, we have more work to do. */ + if (!dsa->stencil[1].enabled) { + /* This just flags that the registers need not be deallocated later */ + *back_fail_reg = fbS_reg; + *back_zfail_reg = fbS_reg; + *back_zpass_reg = fbS_reg; + } + else { + /* Same calculations as above, but for the back stencil */ + if (dsa->stencil[1].fail_op == PIPE_STENCIL_OP_KEEP) { + *back_fail_reg = fbS_reg; + } + else if (dsa->stencil[1].fail_op == dsa->stencil[0].fail_op) { + *back_fail_reg = *fail_reg; + } + else if (dsa->stencil[1].fail_op == zfail_op) { + *back_fail_reg = *zfail_reg; + } + else if (dsa->stencil[1].fail_op == dsa->stencil[0].zpass_op) { + *back_fail_reg = *zpass_reg; + } + else { + *back_fail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[1].fail_op, dsa->stencil[1].ref_value, + 0xff, fbS_reg, *back_fail_reg); + } + + if (back_zfail_op == PIPE_STENCIL_OP_KEEP) { + *back_zfail_reg = fbS_reg; + } + else if (back_zfail_op == dsa->stencil[0].fail_op) { + *back_zfail_reg = *fail_reg; + } + else if (back_zfail_op == zfail_op) { + *back_zfail_reg = *zfail_reg; + } + else if (back_zfail_op == dsa->stencil[0].zpass_op) { + *back_zfail_reg = *zpass_reg; + } + else if (back_zfail_op == dsa->stencil[1].fail_op) { + *back_zfail_reg = *back_fail_reg; + } + else { + *back_zfail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[1].zfail_op, dsa->stencil[1].ref_value, + 0xff, fbS_reg, *back_zfail_reg); + } + + if (dsa->stencil[1].zpass_op == PIPE_STENCIL_OP_KEEP) { + *back_zpass_reg = fbS_reg; + } + else if (dsa->stencil[1].zpass_op == dsa->stencil[0].fail_op) { + *back_zpass_reg = *fail_reg; + } + else if (dsa->stencil[1].zpass_op == zfail_op) { + *back_zpass_reg = *zfail_reg; + } + else if (dsa->stencil[1].zpass_op == dsa->stencil[0].zpass_op) { + *back_zpass_reg = *zpass_reg; + } + else if (dsa->stencil[1].zpass_op == dsa->stencil[1].fail_op) { + *back_zpass_reg = *back_fail_reg; + } + else if (dsa->stencil[1].zpass_op == back_zfail_op) { + *back_zpass_reg = *back_zfail_reg; + } + else { + *back_zfail_reg = spe_allocate_available_register(f); + gen_stencil_values(f, dsa->stencil[1].zpass_op, dsa->stencil[1].ref_value, + 0xff, fbS_reg, *back_zpass_reg); + } + } /* End of calculations for back-facing stencil */ +} + +static boolean +gen_stencil_depth_test(struct spe_function *f, + const struct pipe_depth_stencil_alpha_state *dsa, + const int const facing_reg, + const int mask_reg, const int fragZ_reg, + const int fbZ_reg, const int fbS_reg) +{ + /* True if we've generated code that could require writeback to the + * depth and/or stencil buffers + */ + boolean modified_buffers = false; + + boolean need_to_calculate_stencil_values; + boolean need_to_writemask_stencil_values; + + /* Registers. We may or may not actually allocate these, depending + * on whether the state values indicate that we need them. + */ + unsigned int stencil_pass_reg, stencil_fail_reg; + unsigned int stencil_fail_values, stencil_pass_depth_fail_values, stencil_pass_depth_pass_values; + unsigned int stencil_writemask_reg; + unsigned int zmask_reg; + unsigned int newS_reg; + + /* Stenciling is quite complex: up to six different configurable stencil + * operations/calculations can be required (three each for front-facing + * and back-facing fragments). Many of those operations will likely + * be identical, so there's good reason to try to avoid calculating + * the same values more than once (which unfortunately makes the code less + * straightforward). + * + * To make register management easier, we start a new + * register set; we can release all the registers in the set at + * once, and avoid having to keep track of exactly which registers + * we allocate. We can still allocate and free registers as + * desired (if we know we no longer need a register), but we don't + * have to spend the complexity to track the more difficult variant + * register usage scenarios. + */ + spe_allocate_register_set(f); + + /* Calculate the writemask. If the writemask is trivial (either + * all 0s, meaning that we don't need to calculate any stencil values + * because they're not going to change the stencil anyway, or all 1s, + * meaning that we have to calculate the stencil values but do not + * need to mask them), we can avoid generating code. Don't forget + * that we need to consider backfacing stencil, if enabled. + */ + if (dsa->stencil[0].write_mask == 0x0 && (!dsa->stencil[1].enabled || dsa->stencil[1].write_mask == 0x00)) { + /* Trivial: don't need to calculate stencil values, and don't need to + * write them back to the framebuffer. + */ + need_to_calculate_stencil_values = false; + need_to_writemask_stencil_values = false; + } + else if (dsa->stencil[0].write_mask == 0xff && (!dsa->stencil[1].enabled || dsa->stencil[1].write_mask == 0x00)) { + /* Still trivial, but a little less so. We need to write the stencil + * values, but we don't need to mask them. + */ + need_to_calculate_stencil_values = true; + need_to_writemask_stencil_values = false; + } + else { + /* The general case: calculate, mask, and write */ + need_to_calculate_stencil_values = true; + need_to_writemask_stencil_values = true; + + /* While we're here, generate code that calculates what the + * writemask should be. If backface stenciling is enabled, + * and the backface writemask is not the same as the frontface + * writemask, we'll have to generate code that merges the + * two masks into a single effective mask based on fragment facing. + */ + stencil_writemask_reg = spe_allocate_available_register(f); + spe_load_uint(f, stencil_writemask_reg, dsa->stencil[0].write_mask); + if (dsa->stencil[1].enabled && dsa->stencil[0].write_mask != dsa->stencil[1].write_mask) { + unsigned int back_write_mask_reg = spe_allocate_available_register(f); + spe_load_uint(f, back_write_mask_reg, dsa->stencil[1].write_mask); + spe_selb(f, stencil_writemask_reg, stencil_writemask_reg, back_write_mask_reg, facing_reg); + spe_release_register(f, back_write_mask_reg); + } + } + + /* At least one-sided stenciling must be on. Generate code that + * runs the stencil test on the basic/front-facing stencil, leaving + * the mask of passing stencil bits in stencil_pass_reg. This mask will + * be used both to mask the set of active pixels, and also to + * determine how the stencil buffer changes. + * + * This test will *not* change the value in mask_reg (because we don't + * yet know whether to apply the two-sided stencil or one-sided stencil). + */ + stencil_pass_reg = spe_allocate_available_register(f); + gen_stencil_test(f, &dsa->stencil[0], mask_reg, fbS_reg, stencil_pass_reg); + + /* If two-sided stenciling is on, generate code to run the stencil + * test on the backfacing stencil as well, and combine the two results + * into the one correct result based on facing. + */ + if (dsa->stencil[1].enabled) { + unsigned int temp_reg = spe_allocate_available_register(f); + gen_stencil_test(f, &dsa->stencil[1], mask_reg, fbS_reg, temp_reg); + spe_selb(f, stencil_pass_reg, stencil_pass_reg, temp_reg, facing_reg); + spe_release_register(f, temp_reg); + } + + /* Generate code that, given the mask of valid fragments and the + * mask of valid fragments that passed the stencil test, computes + * the mask of valid fragments that failed the stencil test. We + * have to do this before we run a depth test (because the + * depth test should not be performed on fragments that failed the + * stencil test, and because the depth test will update the + * mask of valid fragments based on the results of the depth test). + */ + stencil_fail_reg = spe_allocate_available_register(f); + spe_andc(f, stencil_fail_reg, mask_reg, stencil_pass_reg); + /* Now remove the stenciled-out pixels from the valid fragment mask, + * so we can later use the valid fragment mask in the depth test. + */ + spe_and(f, mask_reg, mask_reg, stencil_pass_reg); + + /* We may not need to calculate stencil values, if the writemask is off */ + if (need_to_calculate_stencil_values) { + unsigned int back_stencil_fail_values, back_stencil_pass_depth_fail_values, back_stencil_pass_depth_pass_values; + unsigned int front_stencil_fail_values, front_stencil_pass_depth_fail_values, front_stencil_pass_depth_pass_values; + + /* Generate code that calculates exactly which stencil values we need, + * without calculating the same value twice (say, if two different + * stencil ops have the same value). This code will work for one-sided + * and two-sided stenciling (so that we take into account that operations + * may match between front and back stencils), and will also take into + * account whether the depth test is enabled (if the depth test is off, + * we don't need any of the zfail results, because the depth test always + * is considered to pass if it is disabled). Any register value that + * does not need to be calculated will come back with the same value + * that's in fbS_reg. + * + * This function will allocate a variant number of registers that + * will be released as part of the register set. + */ + gen_get_stencil_values(f, dsa, fbS_reg, + &front_stencil_fail_values, &front_stencil_pass_depth_fail_values, + &front_stencil_pass_depth_pass_values, &back_stencil_fail_values, + &back_stencil_pass_depth_fail_values, &back_stencil_pass_depth_pass_values); + + /* Tricky, tricky, tricky - the things we do to create optimal + * code... + * + * The various stencil values registers may overlap with each other + * and with fbS_reg arbitrarily (as any particular operation is + * only calculated once and stored in one register, no matter + * how many times it is used). So we can't change the values + * within those registers directly - if we change a value in a + * register that's being referenced by two different calculations, + * we've just unwittingly changed the second value as well... + * + * Avoid this by allocating new registers to hold the results + * (there may be 2, if the depth test is off, or 3, if it is on). + * These will be released as part of the register set. + */ + if (!dsa->stencil[1].enabled) { + /* The easy case: if two-sided stenciling is *not* enabled, we + * just use the front-sided values. + */ + stencil_fail_values = front_stencil_fail_values; + stencil_pass_depth_fail_values = front_stencil_pass_depth_fail_values; + stencil_pass_depth_pass_values = front_stencil_pass_depth_pass_values; + } + else { /* two-sided stencil enabled */ + /* Allocate new registers for the needed merged values */ + stencil_fail_values = spe_allocate_available_register(f); + spe_selb(f, stencil_fail_values, front_stencil_fail_values, back_stencil_fail_values, facing_reg); + if (dsa->depth.enabled) { + stencil_pass_depth_fail_values = spe_allocate_available_register(f); + spe_selb(f, stencil_pass_depth_fail_values, front_stencil_pass_depth_fail_values, back_stencil_pass_depth_fail_values, facing_reg); + } + else { + stencil_pass_depth_fail_values = fbS_reg; + } + stencil_pass_depth_pass_values = spe_allocate_available_register(f); + spe_selb(f, stencil_pass_depth_pass_values, front_stencil_pass_depth_pass_values, back_stencil_pass_depth_pass_values, facing_reg); + } + } + + /* We now have all the stencil values we need. We also need + * the results of the depth test to figure out which + * stencil values will become the new stencil values. (Even if + * we aren't actually calculating stencil values, we need to apply + * the depth test if it's enabled.) + * + * The code generated by gen_depth_test() returns the results of the + * test in the given register, but also alters the mask_reg based + * on the results of the test. + */ + if (dsa->depth.enabled) { + zmask_reg = spe_allocate_available_register(f); + modified_buffers |= gen_depth_test(f, dsa, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); + } + + if (need_to_calculate_stencil_values) { + /* If we need to writemask the stencil values before going into + * the stencil buffer, we'll have to use a new register to + * hold the new values. If not, we can just keep using the + * current register. + */ + if (need_to_writemask_stencil_values) { + newS_reg = spe_allocate_available_register(f); + spe_move(f, newS_reg, fbS_reg); + modified_buffers = true; + } + else { + newS_reg = fbS_reg; + } + + /* Merge in the selected stencil fail values */ + if (stencil_fail_values != fbS_reg) { + spe_selb(f, newS_reg, newS_reg, stencil_fail_values, stencil_fail_reg); + } + + /* Same for the stencil pass/depth fail values. If this calculation + * is not needed (say, if depth test is off), then the + * stencil_pass_depth_fail_values register will be equal to fbS_reg + * and we'll skip the calculation. + */ + if (stencil_pass_depth_fail_values != fbS_reg) { + /* We don't actually have a stencil pass/depth fail mask yet. + * Calculate it here from the stencil passing mask and the + * depth passing mask. Note that zmask_reg *must* have been + * set above if we're here. + */ + unsigned int stencil_pass_depth_fail_mask = spe_allocate_available_register(f); + spe_andc(f, stencil_pass_depth_fail_mask, stencil_pass_reg, zmask_reg); + + spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_fail_values, stencil_pass_depth_fail_mask); + + spe_release_register(f, stencil_pass_depth_fail_mask); + } + + /* Same for the stencil pass/depth pass mask */ + if (stencil_pass_depth_pass_values != fbS_reg) { + unsigned int stencil_pass_depth_pass_mask = spe_allocate_available_register(f); + spe_and(f, stencil_pass_depth_pass_mask, stencil_pass_reg, zmask_reg); + + spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_pass_values, stencil_pass_depth_pass_mask); + spe_release_register(f, stencil_pass_depth_pass_mask); + } + + /* Almost done. If we need to writemask, do it now, leaving the + * results in the fbS_reg register passed in. If we don't need + * to writemask, then the results are *already* in the fbS_reg, + * so there's nothing more to do. + */ + + if (need_to_writemask_stencil_values) { + /* The Select Bytes command makes a fine writemask. Where + * the mask is 0, the first (original) values are retained, + * effectively masking out changes. Where the mask is 1, the + * second (new) values are retained, incorporating changes. + */ + spe_selb(f, fbS_reg, fbS_reg, newS_reg, stencil_writemask_reg); + } + } /* done calculating stencil values */ + + /* The stencil and/or depth values have been applied, and the + * mask_reg, fbS_reg, and fbZ_reg values have been updated. + * We're all done, except that we've allocated a fair number + * of registers that we didn't bother tracking. Release all + * those registers as part of the register set, and go home. + */ + spe_release_register_set(f); + + /* Return true if we could have modified the stencil and/or + * depth buffers. + */ + return modified_buffers; +} + + /** * Generate SPE code to implement the fragment operations (alpha test, * depth test, stencil test, blending, colormask, and final @@ -1156,6 +1800,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) const int fragB_reg = 10; /* vector float */ const int fragA_reg = 11; /* vector float */ const int mask_reg = 12; /* vector uint */ + const int facing_reg = 13; /* uint */ /* offset of quad from start of tile * XXX assuming 4-byte pixels for color AND Z/stencil!!!! @@ -1183,6 +1828,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_allocate_register(f, fragB_reg); spe_allocate_register(f, fragA_reg); spe_allocate_register(f, mask_reg); + spe_allocate_register(f, facing_reg); quad_offset_reg = spe_allocate_available_register(f); fbRGBA_reg = spe_allocate_available_register(f); @@ -1195,6 +1841,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) ASSERT(TILE_SIZE == 32); + spe_comment(f, 0, "Computing tile location in memory"); spe_rotmi(f, y2_reg, y_reg, -1); /* y2 = y / 2 */ spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */ spe_shli(f, y2_reg, y2_reg, 4); /* y2 *= 16 */ @@ -1205,124 +1852,164 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_release_register(f, y2_reg); } - if (dsa->alpha.enabled) { gen_alpha_test(dsa, f, mask_reg, fragA_reg); } + /* If we need the stencil buffers (because one- or two-sided stencil is + * enabled) or the depth buffer (because the depth test is enabled), + * go grab them. Note that if either one- or two-sided stencil is + * enabled, dsa->stencil[0].enabled will be true. + */ if (dsa->depth.enabled || dsa->stencil[0].enabled) { const enum pipe_format zs_format = cell->framebuffer.zsbuf->format; boolean write_depth_stencil; - int fbZ_reg = spe_allocate_available_register(f); /* Z values */ - int fbS_reg = spe_allocate_available_register(f); /* Stencil values */ + /* We may or may not need to allocate a register for Z or stencil values */ + boolean fbS_reg_set = false, fbZ_reg_set = false; + unsigned int fbS_reg, fbZ_reg = 0; + + spe_comment(f, 0, "Loading Z/stencil tile"); /* fetch quad of depth/stencil values from tile at (x,y) */ /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + /* XXX Not sure this is allowed if we've only got a 16-bit Z buffer... */ spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); - if (dsa->depth.enabled) { - /* Extract Z bits from fbZS_reg into fbZ_reg */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - int mask_reg = spe_allocate_available_register(f); - spe_fsmbi(f, mask_reg, 0x7777); /* mask[0,1,2,3] = 0x00ffffff */ - spe_and(f, fbZ_reg, fbZS_reg, mask_reg); /* fbZ = fbZS & mask */ - spe_release_register(f, mask_reg); - /* OK, fbZ_reg has four 24-bit Z values now */ - } - else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || - zs_format == PIPE_FORMAT_Z24X8_UNORM) { - spe_rotmi(f, fbZ_reg, fbZS_reg, -8); /* fbZ = fbZS >> 8 */ - /* OK, fbZ_reg has four 24-bit Z values now */ - } - else if (zs_format == PIPE_FORMAT_Z32_UNORM) { - spe_move(f, fbZ_reg, fbZS_reg); - /* OK, fbZ_reg has four 32-bit Z values now */ - } - else if (zs_format == PIPE_FORMAT_Z16_UNORM) { - spe_move(f, fbZ_reg, fbZS_reg); - /* OK, fbZ_reg has four 16-bit Z values now */ - } - else { - ASSERT(0); /* invalid format */ - } - - /* Convert fragZ values from float[4] to 16, 24 or 32-bit uint[4] */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM || - zs_format == PIPE_FORMAT_Z24S8_UNORM || - zs_format == PIPE_FORMAT_Z24X8_UNORM) { - /* scale/convert fragZ from float in [0,1] to uint in [0, ~0] */ - spe_cfltu(f, fragZ_reg, fragZ_reg, 32); - /* fragZ = fragZ >> 8 */ - spe_rotmi(f, fragZ_reg, fragZ_reg, -8); - } - else if (zs_format == PIPE_FORMAT_Z32_UNORM) { - /* scale/convert fragZ from float in [0,1] to uint in [0, ~0] */ - spe_cfltu(f, fragZ_reg, fragZ_reg, 32); - } - else if (zs_format == PIPE_FORMAT_Z16_UNORM) { - /* scale/convert fragZ from float in [0,1] to uint in [0, ~0] */ - spe_cfltu(f, fragZ_reg, fragZ_reg, 32); - /* fragZ = fragZ >> 16 */ - spe_rotmi(f, fragZ_reg, fragZ_reg, -16); - } - } - else { - /* no Z test, but set Z to zero so we don't OR-in garbage below */ - spe_load_uint(f, fbZ_reg, 0); /* XXX set to zero for now */ + /* From the Z/stencil buffer format, pull out the bits we need for + * Z and/or stencil. We'll also convert the incoming fragment Z + * value in fragZ_reg from a floating point value in [0.0..1.0] to + * an unsigned integer value with the appropriate resolution. + */ + switch(zs_format) { + + case PIPE_FORMAT_S8Z24_UNORM: /* fall through */ + case PIPE_FORMAT_X8Z24_UNORM: + if (dsa->depth.enabled) { + /* We need the Z part at least */ + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* four 24-bit Z values in the low-order bits */ + spe_and_uint(f, fbZ_reg, fbZS_reg, 0x00ffffff); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 24-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); + } + if (dsa->stencil[0].enabled) { + setup_optional_register(f, &fbS_reg_set, &fbS_reg); + /* four 8-bit Z values in the high-order bits */ + spe_rotmi(f, fbS_reg, fbZS_reg, -24); + } + break; + + case PIPE_FORMAT_Z24S8_UNORM: /* fall through */ + case PIPE_FORMAT_Z24X8_UNORM: + if (dsa->depth.enabled) { + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* shift by 8 to get the upper 24-bit values */ + spe_rotmi(f, fbS_reg, fbZS_reg, -8); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 24-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -8); + } + if (dsa->stencil[0].enabled) { + setup_optional_register(f, &fbS_reg_set, &fbS_reg); + /* 8-bit stencil in the low-order bits - mask them out */ + spe_and_uint(f, fbS_reg, fbZS_reg, 0x000000ff); + } + break; + + case PIPE_FORMAT_Z32_UNORM: + if (dsa->depth.enabled) { + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* Copy over 4 32-bit values */ + spe_move(f, fbZ_reg, fbZS_reg); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 32-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + } + /* No stencil, so can't do anything there */ + break; + + case PIPE_FORMAT_Z16_UNORM: + if (dsa->depth.enabled) { + /* XXX Not sure this is correct, but it was here before, so we're + * going with it for now + */ + setup_optional_register(f, &fbZ_reg_set, &fbZ_reg); + /* Copy over 4 32-bit values */ + spe_move(f, fbZ_reg, fbZS_reg); + + /* Incoming fragZ_reg value is a float in 0.0...1.0; convert + * to a 16-bit unsigned integer + */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 32); + spe_rotmi(f, fragZ_reg, fragZ_reg, -16); + } + /* No stencil */ + break; + + default: + ASSERT(0); /* invalid format */ } - + /* If stencil is enabled, use the stencil-specific code + * generator to generate both the stencil and depth (if needed) + * tests. Otherwise, if only depth is enabled, generate + * a quick depth test. The test generators themselves will + * report back whether the depth/stencil buffer has to be + * written back. + */ if (dsa->stencil[0].enabled) { - /* Extract Stencil bit sfrom fbZS_reg into fbS_reg */ - if (zs_format == PIPE_FORMAT_S8Z24_UNORM || - zs_format == PIPE_FORMAT_X8Z24_UNORM) { - /* XXX extract with a shift */ - ASSERT(0); - } - else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || - zs_format == PIPE_FORMAT_Z24X8_UNORM) { - /* XXX extract with a mask */ - ASSERT(0); - } - } - else { - /* no stencil test, but set to zero so we don't OR-in garbage below */ - spe_load_uint(f, fbS_reg, 0); /* XXX set to zero for now */ - } + /* This will perform the stencil and depth tests, and update + * the mask_reg, fbZ_reg, and fbS_reg as required by the + * tests. + */ + ASSERT(fbS_reg_set); + ASSERT(fbZ_reg_set); + spe_comment(f, 0, "Perform stencil test"); - if (dsa->stencil[0].enabled) { - /* XXX this may involve depth testing too */ - // gen_stencil_test(dsa, f, ... ); - ASSERT(0); + write_depth_stencil = gen_stencil_depth_test(f, dsa, facing_reg, mask_reg, fragZ_reg, fbZ_reg, fbS_reg); } else if (dsa->depth.enabled) { int zmask_reg = spe_allocate_available_register(f); - gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); + spe_comment(f, 0, "Perform depth test"); + write_depth_stencil = gen_depth_test(f, dsa, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); spe_release_register(f, zmask_reg); } - - /* do we need to write Z and/or Stencil back into framebuffer? */ - write_depth_stencil = (dsa->depth.writemask | - dsa->stencil[0].write_mask | - dsa->stencil[1].write_mask); + else { + write_depth_stencil = false; + } if (write_depth_stencil) { /* Merge latest Z and Stencil values into fbZS_reg. * fbZ_reg has four Z vals in bits [23..0] or bits [15..0]. * fbS_reg has four 8-bit Z values in bits [7..0]. */ + spe_comment(f, 0, "Storing depth/stencil values"); if (zs_format == PIPE_FORMAT_S8Z24_UNORM || zs_format == PIPE_FORMAT_X8Z24_UNORM) { - spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ - spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + if (fbS_reg_set) { + spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } + else { + spe_move(f, fbZS_reg, fbZ_reg); + } } else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || zs_format == PIPE_FORMAT_Z24X8_UNORM) { spe_shli(f, fbZ_reg, fbZ_reg, 8); /* fbZ = fbZ << 8 */ - spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + if (fbS_reg_set) { + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } } else if (zs_format == PIPE_FORMAT_Z32_UNORM) { spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */ @@ -1341,11 +2028,10 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); } - spe_release_register(f, fbZ_reg); - spe_release_register(f, fbS_reg); + release_optional_register(f, &fbZ_reg_set, fbZ_reg); + release_optional_register(f, &fbS_reg_set, fbS_reg); } - /* Get framebuffer quad/colors. We'll need these for blending, * color masking, and to obey the quad/pixel mask. * Load: fbRGBA_reg = memory[color_tile + quad_offset] @@ -1354,8 +2040,8 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) */ spe_lqx(f, fbRGBA_reg, color_tile_reg, quad_offset_reg); - if (blend->blend_enable) { + spe_comment(f, 0, "Perform blending"); gen_blend(blend, blend_color, f, color_format, fragR_reg, fragG_reg, fragB_reg, fragA_reg, fbRGBA_reg); } @@ -1369,19 +2055,21 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) int rgba_reg = spe_allocate_available_register(f); /* Pack four float colors as four 32-bit int colors */ + spe_comment(f, 0, "Convert fragment colors to framebuffer colors"); gen_pack_colors(f, color_format, fragR_reg, fragG_reg, fragB_reg, fragA_reg, rgba_reg); if (blend->logicop_enable) { + spe_comment(f, 0, "Compute logic op"); gen_logicop(blend, f, rgba_reg, fbRGBA_reg); } if (blend->colormask != PIPE_MASK_RGBA) { + spe_comment(f, 0, "Compute color mask"); gen_colormask(f, blend->colormask, color_format, rgba_reg, fbRGBA_reg); } - /* Mix fragment colors with framebuffer colors using the quad/pixel mask: * if (mask[i]) * rgba[i] = rgba[i]; @@ -1393,6 +2081,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) /* Store updated quad in tile: * memory[color_tile + quad_offset] = rgba_reg; */ + spe_comment(f, 0, "Store framebuffer colors"); spe_stqx(f, rgba_reg, color_tile_reg, quad_offset_reg); spe_release_register(f, rgba_reg); diff --git a/src/gallium/drivers/cell/ppu/cell_render.c b/src/gallium/drivers/cell/ppu/cell_render.c index dd25ae880e..79cb8df82f 100644 --- a/src/gallium/drivers/cell/ppu/cell_render.c +++ b/src/gallium/drivers/cell/ppu/cell_render.c @@ -152,6 +152,7 @@ cell_flush_prim_buffer(struct cell_context *cell) struct cell_command_render *render = &cell_global.command[i].render; render->prim_type = PIPE_PRIM_TRIANGLES; render->num_verts = cell->prim_buffer.num_verts; + render->front_winding = cell->rasterizer->front_winding; render->vertex_size = cell->vertex_info->size * 4; render->xmin = cell->prim_buffer.xmin; render->ymin = cell->prim_buffer.ymin; diff --git a/src/gallium/drivers/cell/ppu/cell_vbuf.c b/src/gallium/drivers/cell/ppu/cell_vbuf.c index aa63435b93..578ddf62dc 100644 --- a/src/gallium/drivers/cell/ppu/cell_vbuf.c +++ b/src/gallium/drivers/cell/ppu/cell_vbuf.c @@ -214,6 +214,7 @@ cell_vbuf_draw(struct vbuf_render *vbr, render->opcode = CELL_CMD_RENDER; render->prim_type = cvbr->prim; + render->front_winding = cell->rasterizer->front_winding; render->num_indexes = nr_indices; render->min_index = min_index; diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h index 29a305232e..1cd577c23c 100644 --- a/src/gallium/drivers/cell/spu/spu_main.h +++ b/src/gallium/drivers/cell/spu/spu_main.h @@ -73,7 +73,8 @@ typedef void (*spu_fragment_ops_func)(uint x, uint y, vector float fragGreen, vector float fragBlue, vector float fragAlpha, - vector unsigned int mask); + vector unsigned int mask, + uint facing); /** Function for running fragment program */ typedef void (*spu_fragment_program_func)(vector float *inputs, diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c index f107764fb2..d252fa6dc1 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c @@ -57,7 +57,8 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fragG, vector float fragB, vector float fragA, - vector unsigned int mask) + vector unsigned int mask, + uint facing) { vector float frag_aos[4]; unsigned int fbc0, fbc1, fbc2, fbc3 ; /* framebuffer/tile colors */ @@ -433,23 +434,23 @@ spu_fallback_fragment_ops(uint x, uint y, /* Form bitmask depending on color buffer format and colormask bits */ switch (spu.fb.color_format) { case PIPE_FORMAT_A8R8G8B8_UNORM: - if (spu.blend.colormask & (1<<0)) + if (spu.blend.colormask & PIPE_MASK_R) cmask |= 0x00ff0000; /* red */ - if (spu.blend.colormask & (1<<1)) + if (spu.blend.colormask & PIPE_MASK_G) cmask |= 0x0000ff00; /* green */ - if (spu.blend.colormask & (1<<2)) + if (spu.blend.colormask & PIPE_MASK_B) cmask |= 0x000000ff; /* blue */ - if (spu.blend.colormask & (1<<3)) + if (spu.blend.colormask & PIPE_MASK_A) cmask |= 0xff000000; /* alpha */ break; case PIPE_FORMAT_B8G8R8A8_UNORM: - if (spu.blend.colormask & (1<<0)) + if (spu.blend.colormask & PIPE_MASK_R) cmask |= 0x0000ff00; /* red */ - if (spu.blend.colormask & (1<<1)) + if (spu.blend.colormask & PIPE_MASK_G) cmask |= 0x00ff0000; /* green */ - if (spu.blend.colormask & (1<<2)) + if (spu.blend.colormask & PIPE_MASK_B) cmask |= 0xff000000; /* blue */ - if (spu.blend.colormask & (1<<3)) + if (spu.blend.colormask & PIPE_MASK_A) cmask |= 0x000000ff; /* alpha */ break; default: diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h index f817abf046..a61689c83a 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h @@ -38,7 +38,8 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fragGreen, vector float fragBlue, vector float fragAlpha, - vector unsigned int mask); + vector unsigned int mask, + uint facing); #endif /* SPU_PER_FRAGMENT_OP */ diff --git a/src/gallium/drivers/cell/spu/spu_render.c b/src/gallium/drivers/cell/spu/spu_render.c index 305dc98881..82dbeb26b7 100644 --- a/src/gallium/drivers/cell/spu/spu_render.c +++ b/src/gallium/drivers/cell/spu/spu_render.c @@ -279,7 +279,7 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) v1 = (const float *) (vertices + indexes[j+1] * vertex_size); v2 = (const float *) (vertices + indexes[j+2] * vertex_size); - drawn += tri_draw(v0, v1, v2, tx, ty); + drawn += tri_draw(v0, v1, v2, tx, ty, render->front_winding); } //printf("SPU %u: drew %u of %u\n", spu.init.id, drawn, render->num_indexes/3); @@ -297,5 +297,3 @@ cmd_render(const struct cell_command_render *render, uint *pos_incr) printf("SPU %u: RENDER done\n", spu.init.id); } - - diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index 0a8fb56a62..6039cd80b2 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -118,6 +118,8 @@ struct setup_stage { float oneoverarea; + uint facing; + uint tx, ty; int cliprect_minx, cliprect_maxx, cliprect_miny, cliprect_maxy; @@ -274,7 +276,7 @@ eval_z(float x, float y) * overall. */ static INLINE void -emit_quad( int x, int y, mask_t mask ) +emit_quad( int x, int y, mask_t mask) { /* If any bits in mask are set... */ if (spu_extract(spu_orx(mask), 0)) { @@ -344,7 +346,8 @@ emit_quad( int x, int y, mask_t mask ) fragZ, soa_frag[0], soa_frag[1], soa_frag[2], soa_frag[3], - mask); + mask, + setup.facing); } } @@ -379,7 +382,8 @@ emit_quad( int x, int y, mask_t mask ) outputs[0*4+1], outputs[0*4+2], outputs[0*4+3], - mask); + mask, + setup.facing); } } } @@ -483,7 +487,7 @@ static void flush_spans( void ) */ for (x = block(minleft); x <= block(maxright); x += 2) { #if 1 - emit_quad( x, setup.span.y, calculate_mask( x ) ); + emit_quad( x, setup.span.y, calculate_mask( x )); #endif } @@ -902,13 +906,28 @@ static void subtriangle( struct edge *eleft, eright->sy += lines; } +static float +determinant( const float *v0, + const float *v1, + const float *v2 ) +{ + /* edge vectors e = v0 - v2, f = v1 - v2 */ + const float ex = v0[0] - v2[0]; + const float ey = v0[1] - v2[1]; + const float fx = v1[0] - v2[0]; + const float fy = v1[1] - v2[1]; + + /* det = cross(e,f).z */ + return ex * fy - ey * fx; +} + /** * Draw triangle into tile at (tx, ty) (tile coords) * The tile data should have already been fetched. */ boolean -tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty) +tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty, uint front_winding) { setup.tx = tx; setup.ty = ty; @@ -919,6 +938,12 @@ tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty) setup.cliprect_maxx = (tx + 1) * TILE_SIZE; setup.cliprect_maxy = (ty + 1) * TILE_SIZE; + /* Before we sort vertices, determine the facing of the triangle, + * which will be needed for front/back-face stencil application + */ + float det = determinant(v0, v1, v2); + setup.facing = (det > 0.0) ^ (front_winding == PIPE_WINDING_CW); + if (!setup_sort_vertices((struct vertex_header *) v0, (struct vertex_header *) v1, (struct vertex_header *) v2)) { diff --git a/src/gallium/drivers/cell/spu/spu_tri.h b/src/gallium/drivers/cell/spu/spu_tri.h index aa694dd7c9..abc3d35160 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.h +++ b/src/gallium/drivers/cell/spu/spu_tri.h @@ -31,7 +31,7 @@ extern boolean -tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty); +tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty, uint front_winding); #endif /* SPU_TRI_H */ -- cgit v1.2.3 From 90027f85786406133a5180998a75fb612b6a221e Mon Sep 17 00:00:00 2001 From: Robert Ellison Date: Tue, 11 Nov 2008 13:57:10 -0700 Subject: CELL: two-sided stencil fixes With these changes, the tests/stencil_twoside test now works. - Eliminate blending from the stencil_twoside test, as it produces an unneeded dependency on having blending working - The spe_splat() function will now work if the register being splatted and the destination register are the same - Separate fragment code generated for front-facing and back-facing fragments. Often these are the same; if two-sided stenciling is on, they can be different. This is easier and faster than generating code that does both tests and merges the results. - Fixed a cut/paste bug where if the back Z-pass stencil operation were different from all the other operations, the back Z-fail results were incorrect. --- progs/tests/stencil_twoside.c | 2 - src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c | 7 +- src/gallium/drivers/cell/common.h | 6 +- src/gallium/drivers/cell/ppu/cell_gen_fragment.c | 239 ++++++--------------- src/gallium/drivers/cell/ppu/cell_gen_fragment.h | 2 +- src/gallium/drivers/cell/ppu/cell_state_emit.c | 19 +- src/gallium/drivers/cell/spu/spu_command.c | 6 +- src/gallium/drivers/cell/spu/spu_main.c | 6 +- src/gallium/drivers/cell/spu/spu_main.h | 10 +- src/gallium/drivers/cell/spu/spu_per_fragment_op.c | 3 +- src/gallium/drivers/cell/spu/spu_per_fragment_op.h | 3 +- src/gallium/drivers/cell/spu/spu_tri.c | 20 +- 12 files changed, 115 insertions(+), 208 deletions(-) (limited to 'src/gallium/drivers/cell/spu/spu_per_fragment_op.h') diff --git a/progs/tests/stencil_twoside.c b/progs/tests/stencil_twoside.c index be9d9a776a..8826c46fc2 100644 --- a/progs/tests/stencil_twoside.c +++ b/progs/tests/stencil_twoside.c @@ -115,7 +115,6 @@ static void Display( void ) glVertex2f(-1, 1); glEnd(); - if (use20syntax) { stencil_func_separate(GL_FRONT, GL_ALWAYS, 0, ~0); stencil_func_separate(GL_BACK, GL_ALWAYS, 0, ~0); @@ -279,7 +278,6 @@ static void Init( void ) stencil_op_separate = glutGetProcAddress( "glStencilOpSeparate" ); printf("\nAll 5 squares should be the same color.\n"); - glEnable( GL_BLEND ); } diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c index f8568f690b..1bd9f1c8dd 100644 --- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c +++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c @@ -958,9 +958,12 @@ spe_compare_greater_uint(struct spe_function *p, unsigned rT, unsigned rA, unsig void spe_splat(struct spe_function *p, unsigned rT, unsigned rA) { + /* Use a temporary, just in case rT == rA */ + unsigned int tmp_reg = spe_allocate_available_register(p); /* Duplicate bytes 0, 1, 2, and 3 across the whole register */ - spe_ila(p, rT, 0x00010203); - spe_shufb(p, rT, rA, rA, rT); + spe_ila(p, tmp_reg, 0x00010203); + spe_shufb(p, rT, rA, rA, tmp_reg); + spe_release_register(p, tmp_reg); } diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h index 87488ea2d7..a670ed3c6e 100644 --- a/src/gallium/drivers/cell/common.h +++ b/src/gallium/drivers/cell/common.h @@ -130,6 +130,9 @@ #define CELL_FENCE_EMITTED 1 #define CELL_FENCE_SIGNALLED 2 +#define CELL_FACING_FRONT 0 +#define CELL_FACING_BACK 1 + struct cell_fence { /** There's a 16-byte status qword per SPU */ @@ -160,7 +163,8 @@ struct cell_command_fragment_ops struct pipe_depth_stencil_alpha_state dsa; struct pipe_blend_state blend; struct pipe_blend_color blend_color; - unsigned code[SPU_MAX_FRAGMENT_OPS_INSTS]; + unsigned code_front[SPU_MAX_FRAGMENT_OPS_INSTS]; + unsigned code_back[SPU_MAX_FRAGMENT_OPS_INSTS]; }; diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c index d9c3ff3f4d..6e425eafaa 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -1412,144 +1412,72 @@ gen_stencil_values(struct spe_function *f, unsigned int stencil_op, * and released by the corresponding spe_release_register_set() call. */ static void -gen_get_stencil_values(struct spe_function *f, const struct pipe_depth_stencil_alpha_state *dsa, +gen_get_stencil_values(struct spe_function *f, const struct pipe_stencil_state *stencil, + const unsigned int depth_enabled, unsigned int fbS_reg, unsigned int *fail_reg, unsigned int *zfail_reg, - unsigned int *zpass_reg, unsigned int *back_fail_reg, - unsigned int *back_zfail_reg, unsigned int *back_zpass_reg) + unsigned int *zpass_reg) { - unsigned zfail_op, back_zfail_op; + unsigned zfail_op; /* Stenciling had better be enabled here */ - ASSERT(dsa->stencil[0].enabled); + ASSERT(stencil->enabled); /* If the depth test is not enabled, it is treated as though it always - * passes. In particular, that means that the "zfail_op" (and the backfacing - * counterpart, if active) are not considered - a failing stencil test will - * trigger the "fail_op", and a passing stencil test will trigger the - * "zpass_op". + * passes, which means that the zfail_op is not considered - a + * failing stencil test triggers the fail_op, and a passing one + * triggers the zpass_op * - * By overriding the operations in this case to be PIPE_STENCIL_OP_KEEP, - * we keep them from being calculated. + * As an optimization, override calculation of the zfail_op values + * if they aren't going to be used. By setting the value of + * the operation to PIPE_STENCIL_OP_KEEP, its value will be assumed + * to match the incoming stencil values, and no calculation will + * be done. */ - if (dsa->depth.enabled) { - zfail_op = dsa->stencil[0].zfail_op; - back_zfail_op = dsa->stencil[1].zfail_op; + if (depth_enabled) { + zfail_op = stencil->zfail_op; } else { zfail_op = PIPE_STENCIL_OP_KEEP; - back_zfail_op = PIPE_STENCIL_OP_KEEP; } /* One-sided or front-facing stencil */ - if (dsa->stencil[0].fail_op == PIPE_STENCIL_OP_KEEP) { + if (stencil->fail_op == PIPE_STENCIL_OP_KEEP) { *fail_reg = fbS_reg; } else { *fail_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[0].fail_op, dsa->stencil[0].ref_value, + gen_stencil_values(f, stencil->fail_op, stencil->ref_value, 0xff, fbS_reg, *fail_reg); } + /* Check the possibly overridden value, not the structure value */ if (zfail_op == PIPE_STENCIL_OP_KEEP) { *zfail_reg = fbS_reg; } - else if (zfail_op == dsa->stencil[0].fail_op) { + else if (zfail_op == stencil->fail_op) { *zfail_reg = *fail_reg; } else { *zfail_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[0].zfail_op, dsa->stencil[0].ref_value, + gen_stencil_values(f, stencil->zfail_op, stencil->ref_value, 0xff, fbS_reg, *zfail_reg); } - if (dsa->stencil[0].zpass_op == PIPE_STENCIL_OP_KEEP) { + if (stencil->zpass_op == PIPE_STENCIL_OP_KEEP) { *zpass_reg = fbS_reg; } - else if (dsa->stencil[0].zpass_op == dsa->stencil[0].fail_op) { + else if (stencil->zpass_op == stencil->fail_op) { *zpass_reg = *fail_reg; } - else if (dsa->stencil[0].zpass_op == zfail_op) { + else if (stencil->zpass_op == zfail_op) { *zpass_reg = *zfail_reg; } else { *zpass_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[0].zpass_op, dsa->stencil[0].ref_value, + gen_stencil_values(f, stencil->zpass_op, stencil->ref_value, 0xff, fbS_reg, *zpass_reg); } - - /* If two-sided stencil is enabled, we have more work to do. */ - if (!dsa->stencil[1].enabled) { - /* This just flags that the registers need not be deallocated later */ - *back_fail_reg = fbS_reg; - *back_zfail_reg = fbS_reg; - *back_zpass_reg = fbS_reg; - } - else { - /* Same calculations as above, but for the back stencil */ - if (dsa->stencil[1].fail_op == PIPE_STENCIL_OP_KEEP) { - *back_fail_reg = fbS_reg; - } - else if (dsa->stencil[1].fail_op == dsa->stencil[0].fail_op) { - *back_fail_reg = *fail_reg; - } - else if (dsa->stencil[1].fail_op == zfail_op) { - *back_fail_reg = *zfail_reg; - } - else if (dsa->stencil[1].fail_op == dsa->stencil[0].zpass_op) { - *back_fail_reg = *zpass_reg; - } - else { - *back_fail_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[1].fail_op, dsa->stencil[1].ref_value, - 0xff, fbS_reg, *back_fail_reg); - } - - if (back_zfail_op == PIPE_STENCIL_OP_KEEP) { - *back_zfail_reg = fbS_reg; - } - else if (back_zfail_op == dsa->stencil[0].fail_op) { - *back_zfail_reg = *fail_reg; - } - else if (back_zfail_op == zfail_op) { - *back_zfail_reg = *zfail_reg; - } - else if (back_zfail_op == dsa->stencil[0].zpass_op) { - *back_zfail_reg = *zpass_reg; - } - else if (back_zfail_op == dsa->stencil[1].fail_op) { - *back_zfail_reg = *back_fail_reg; - } - else { - *back_zfail_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[1].zfail_op, dsa->stencil[1].ref_value, - 0xff, fbS_reg, *back_zfail_reg); - } - - if (dsa->stencil[1].zpass_op == PIPE_STENCIL_OP_KEEP) { - *back_zpass_reg = fbS_reg; - } - else if (dsa->stencil[1].zpass_op == dsa->stencil[0].fail_op) { - *back_zpass_reg = *fail_reg; - } - else if (dsa->stencil[1].zpass_op == zfail_op) { - *back_zpass_reg = *zfail_reg; - } - else if (dsa->stencil[1].zpass_op == dsa->stencil[0].zpass_op) { - *back_zpass_reg = *zpass_reg; - } - else if (dsa->stencil[1].zpass_op == dsa->stencil[1].fail_op) { - *back_zpass_reg = *back_fail_reg; - } - else if (dsa->stencil[1].zpass_op == back_zfail_op) { - *back_zpass_reg = *back_zfail_reg; - } - else { - *back_zfail_reg = spe_allocate_available_register(f); - gen_stencil_values(f, dsa->stencil[1].zpass_op, dsa->stencil[1].ref_value, - 0xff, fbS_reg, *back_zpass_reg); - } - } /* End of calculations for back-facing stencil */ } /* Note that fbZ_reg may *not* be set on entry, if in fact @@ -1559,7 +1487,7 @@ gen_get_stencil_values(struct spe_function *f, const struct pipe_depth_stencil_a static boolean gen_stencil_depth_test(struct spe_function *f, const struct pipe_depth_stencil_alpha_state *dsa, - const int const facing_reg, + const uint facing, const int mask_reg, const int fragZ_reg, const int fbZ_reg, const int fbS_reg) { @@ -1571,6 +1499,8 @@ gen_stencil_depth_test(struct spe_function *f, boolean need_to_calculate_stencil_values; boolean need_to_writemask_stencil_values; + struct pipe_stencil_state *stencil; + /* Registers. We may or may not actually allocate these, depending * on whether the state values indicate that we need them. */ @@ -1598,6 +1528,20 @@ gen_stencil_depth_test(struct spe_function *f, spe_comment(f, 0, "Allocating stencil register set"); spe_allocate_register_set(f); + /* The facing we're given is the fragment facing; it doesn't + * exactly match the stencil facing. If stencil is enabled, + * but two-sided stencil is *not* enabled, we use the same + * stencil settings for both front- and back-facing fragments. + * We only use the "back-facing" stencil for backfacing fragments + * if two-sided stenciling is enabled. + */ + if (facing == CELL_FACING_BACK && dsa->stencil[1].enabled) { + stencil = &dsa->stencil[1]; + } + else { + stencil = &dsa->stencil[0]; + } + /* Calculate the writemask. If the writemask is trivial (either * all 0s, meaning that we don't need to calculate any stencil values * because they're not going to change the stencil anyway, or all 1s, @@ -1608,24 +1552,20 @@ gen_stencil_depth_test(struct spe_function *f, * Note that if the backface stencil is *not* enabled, the backface * stencil will have the same values as the frontface stencil. */ - if (dsa->stencil[0].fail_op == PIPE_STENCIL_OP_KEEP && - dsa->stencil[0].zfail_op == PIPE_STENCIL_OP_KEEP && - dsa->stencil[0].zpass_op == PIPE_STENCIL_OP_KEEP && - dsa->stencil[1].fail_op == PIPE_STENCIL_OP_KEEP && - dsa->stencil[1].zfail_op == PIPE_STENCIL_OP_KEEP && - dsa->stencil[1].zpass_op == PIPE_STENCIL_OP_KEEP) { - /* No changes to any stencil values */ + if (stencil->fail_op == PIPE_STENCIL_OP_KEEP && + stencil->zfail_op == PIPE_STENCIL_OP_KEEP && + stencil->zpass_op == PIPE_STENCIL_OP_KEEP) { need_to_calculate_stencil_values = false; need_to_writemask_stencil_values = false; } - else if (dsa->stencil[0].write_mask == 0x0 && dsa->stencil[1].write_mask == 0x0) { + else if (stencil->write_mask == 0x0) { /* All changes are writemasked out, so no need to calculate * what those changes might be, and no need to write anything back. */ need_to_calculate_stencil_values = false; need_to_writemask_stencil_values = false; } - else if (dsa->stencil[0].write_mask == 0xff && dsa->stencil[1].write_mask == 0xff) { + else if (stencil->write_mask == 0xff) { /* Still trivial, but a little less so. We need to write the stencil * values, but we don't need to mask them. */ @@ -1645,14 +1585,7 @@ gen_stencil_depth_test(struct spe_function *f, */ spe_comment(f, 0, "Computing stencil writemask"); stencil_writemask_reg = spe_allocate_available_register(f); - spe_load_uint(f, stencil_writemask_reg, dsa->stencil[0].write_mask); - if (dsa->stencil[1].enabled && dsa->stencil[0].write_mask != dsa->stencil[1].write_mask) { - unsigned int back_write_mask_reg = spe_allocate_available_register(f); - spe_comment(f, 0, "Resolving two-sided stencil writemask"); - spe_load_uint(f, back_write_mask_reg, dsa->stencil[1].write_mask); - spe_selb(f, stencil_writemask_reg, stencil_writemask_reg, back_write_mask_reg, facing_reg); - spe_release_register(f, back_write_mask_reg); - } + spe_load_uint(f, stencil_writemask_reg, dsa->stencil[facing].write_mask); } /* At least one-sided stenciling must be on. Generate code that @@ -1666,19 +1599,7 @@ gen_stencil_depth_test(struct spe_function *f, */ spe_comment(f, 0, "Running basic stencil test"); stencil_pass_reg = spe_allocate_available_register(f); - gen_stencil_test(f, &dsa->stencil[0], 0xff, mask_reg, fbS_reg, stencil_pass_reg); - - /* If two-sided stenciling is on, generate code to run the stencil - * test on the backfacing stencil as well, and combine the two results - * into the one correct result based on facing. - */ - if (dsa->stencil[1].enabled) { - unsigned int temp_reg = spe_allocate_available_register(f); - spe_comment(f, 0, "Running backface stencil test"); - gen_stencil_test(f, &dsa->stencil[1], 0xff, mask_reg, fbS_reg, temp_reg); - spe_selb(f, stencil_pass_reg, stencil_pass_reg, temp_reg, facing_reg); - spe_release_register(f, temp_reg); - } + gen_stencil_test(f, stencil, 0xff, mask_reg, fbS_reg, stencil_pass_reg); /* Generate code that, given the mask of valid fragments and the * mask of valid fragments that passed the stencil test, computes @@ -1698,9 +1619,6 @@ gen_stencil_depth_test(struct spe_function *f, /* We may not need to calculate stencil values, if the writemask is off */ if (need_to_calculate_stencil_values) { - unsigned int back_stencil_fail_values, back_stencil_pass_depth_fail_values, back_stencil_pass_depth_pass_values; - unsigned int front_stencil_fail_values, front_stencil_pass_depth_fail_values, front_stencil_pass_depth_pass_values; - /* Generate code that calculates exactly which stencil values we need, * without calculating the same value twice (say, if two different * stencil ops have the same value). This code will work for one-sided @@ -1715,51 +1633,11 @@ gen_stencil_depth_test(struct spe_function *f, * This function will allocate a variant number of registers that * will be released as part of the register set. */ - spe_comment(f, 0, "Computing stencil values"); - gen_get_stencil_values(f, dsa, fbS_reg, - &front_stencil_fail_values, &front_stencil_pass_depth_fail_values, - &front_stencil_pass_depth_pass_values, &back_stencil_fail_values, - &back_stencil_pass_depth_fail_values, &back_stencil_pass_depth_pass_values); - - /* Tricky, tricky, tricky - the things we do to create optimal - * code... - * - * The various stencil values registers may overlap with each other - * and with fbS_reg arbitrarily (as any particular operation is - * only calculated once and stored in one register, no matter - * how many times it is used). So we can't change the values - * within those registers directly - if we change a value in a - * register that's being referenced by two different calculations, - * we've just unwittingly changed the second value as well... - * - * Avoid this by allocating new registers to hold the results - * (there may be 2, if the depth test is off, or 3, if it is on). - * These will be released as part of the register set. - */ - if (!dsa->stencil[1].enabled) { - /* The easy case: if two-sided stenciling is *not* enabled, we - * just use the front-sided values. - */ - stencil_fail_values = front_stencil_fail_values; - stencil_pass_depth_fail_values = front_stencil_pass_depth_fail_values; - stencil_pass_depth_pass_values = front_stencil_pass_depth_pass_values; - } - else { /* two-sided stencil enabled */ - spe_comment(f, 0, "Resolving backface stencil values"); - /* Allocate new registers for the needed merged values */ - stencil_fail_values = spe_allocate_available_register(f); - spe_selb(f, stencil_fail_values, front_stencil_fail_values, back_stencil_fail_values, facing_reg); - if (dsa->depth.enabled) { - stencil_pass_depth_fail_values = spe_allocate_available_register(f); - spe_selb(f, stencil_pass_depth_fail_values, front_stencil_pass_depth_fail_values, back_stencil_pass_depth_fail_values, facing_reg); - } - else { - stencil_pass_depth_fail_values = fbS_reg; - } - stencil_pass_depth_pass_values = spe_allocate_available_register(f); - spe_selb(f, stencil_pass_depth_pass_values, front_stencil_pass_depth_pass_values, back_stencil_pass_depth_pass_values, facing_reg); - } - } + spe_comment(f, 0, facing == CELL_FACING_FRONT ? "Computing front-facing stencil values" : "Computing back-facing stencil values"); + gen_get_stencil_values(f, stencil, dsa->depth.enabled, fbS_reg, + &stencil_fail_values, &stencil_pass_depth_fail_values, + &stencil_pass_depth_pass_values); + } /* We now have all the stencil values we need. We also need * the results of the depth test to figure out which @@ -1896,10 +1774,12 @@ gen_stencil_depth_test(struct spe_function *f, * should be much faster. * * \param cell the rendering context (in) + * \param facing whether the generated code is for front-facing or + * back-facing fragments * \param f the generated function (out) */ void -cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) +cell_gen_fragment_function(struct cell_context *cell, uint facing, struct spe_function *f) { const struct pipe_depth_stencil_alpha_state *dsa = cell->depth_stencil; const struct pipe_blend_state *blend = cell->blend; @@ -1917,7 +1797,8 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) const int fragB_reg = 10; /* vector float */ const int fragA_reg = 11; /* vector float */ const int mask_reg = 12; /* vector uint */ - const int facing_reg = 13; /* uint */ + + ASSERT(facing == CELL_FACING_FRONT || facing == CELL_FACING_BACK); /* offset of quad from start of tile * XXX assuming 4-byte pixels for color AND Z/stencil!!!! @@ -1945,7 +1826,6 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_allocate_register(f, fragB_reg); spe_allocate_register(f, fragA_reg); spe_allocate_register(f, mask_reg); - spe_allocate_register(f, facing_reg); quad_offset_reg = spe_allocate_available_register(f); fbRGBA_reg = spe_allocate_available_register(f); @@ -1969,6 +1849,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) spe_release_register(f, y2_reg); } + /* Generate the alpha test, if needed. */ if (dsa->alpha.enabled) { gen_alpha_test(dsa, f, mask_reg, fragA_reg); } @@ -2095,7 +1976,7 @@ cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) * gen_stencil_depth_test() function must ignore the * fbZ_reg register if depth is not enabled. */ - write_depth_stencil = gen_stencil_depth_test(f, dsa, facing_reg, mask_reg, fragZ_reg, fbZ_reg, fbS_reg); + write_depth_stencil = gen_stencil_depth_test(f, dsa, facing, mask_reg, fragZ_reg, fbZ_reg, fbS_reg); } else if (dsa->depth.enabled) { int zmask_reg = spe_allocate_available_register(f); diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h index b59de198dc..2fabfdfb08 100644 --- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h @@ -31,7 +31,7 @@ extern void -cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f); +cell_gen_fragment_function(struct cell_context *cell, uint facing, struct spe_function *f); #endif /* CELL_GEN_FRAGMENT_H */ diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c index dd2d7f7d1e..031b27f11f 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -75,23 +75,29 @@ lookup_fragment_ops(struct cell_context *cell) * If not found, create/save new fragment ops command. */ if (!ops) { - struct spe_function spe_code; + struct spe_function spe_code_front, spe_code_back; if (0) debug_printf("**** Create New Fragment Ops\n"); /* Prepare the buffer that will hold the generated code. */ - spe_init_func(&spe_code, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + spe_init_func(&spe_code_front, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + spe_init_func(&spe_code_back, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); - /* generate new code */ - cell_gen_fragment_function(cell, &spe_code); + /* generate new code. Always generate new code for both front-facing + * and back-facing fragments, even if it's the same code in both + * cases. + */ + cell_gen_fragment_function(cell, CELL_FACING_FRONT, &spe_code_front); + cell_gen_fragment_function(cell, CELL_FACING_BACK, &spe_code_back); /* alloc new fragment ops command */ ops = CALLOC_STRUCT(cell_command_fragment_ops); /* populate the new cell_command_fragment_ops object */ ops->opcode = CELL_CMD_STATE_FRAGMENT_OPS; - memcpy(ops->code, spe_code.store, spe_code_size(&spe_code)); + memcpy(ops->code_front, spe_code_front.store, spe_code_size(&spe_code_front)); + memcpy(ops->code_back, spe_code_back.store, spe_code_size(&spe_code_back)); ops->dsa = *cell->depth_stencil; ops->blend = *cell->blend; @@ -99,7 +105,8 @@ lookup_fragment_ops(struct cell_context *cell) util_keymap_insert(cell->fragment_ops_cache, &key, ops, NULL); /* release rtasm buffer */ - spe_release_func(&spe_code); + spe_release_func(&spe_code_front); + spe_release_func(&spe_code_back); } else { if (0) diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c index d726622d94..d5faf4e3aa 100644 --- a/src/gallium/drivers/cell/spu/spu_command.c +++ b/src/gallium/drivers/cell/spu/spu_command.c @@ -214,7 +214,8 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n"); /* Copy SPU code from batch buffer to spu buffer */ - memcpy(spu.fragment_ops_code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4); + memcpy(spu.fragment_ops_code_front, fops->code_front, SPU_MAX_FRAGMENT_OPS_INSTS * 4); + memcpy(spu.fragment_ops_code_back, fops->code_back, SPU_MAX_FRAGMENT_OPS_INSTS * 4); /* Copy state info (for fallback case only) */ memcpy(&spu.depth_stencil_alpha, &fops->dsa, sizeof(fops->dsa)); memcpy(&spu.blend, &fops->blend, sizeof(fops->blend)); @@ -234,7 +235,8 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) * raw state records that the fallback code requires. */ if ((spu.init.debug_flags & CELL_DEBUG_FRAGMENT_OP_FALLBACK) == 0) { - spu.fragment_ops = (spu_fragment_ops_func) spu.fragment_ops_code; + spu.fragment_ops[CELL_FACING_FRONT] = (spu_fragment_ops_func) spu.fragment_ops_code_front; + spu.fragment_ops[CELL_FACING_BACK] = (spu_fragment_ops_func) spu.fragment_ops_code_back; } else { /* otherwise, the default fallback code remains in place */ diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c index c8bb251905..7033f6037d 100644 --- a/src/gallium/drivers/cell/spu/spu_main.c +++ b/src/gallium/drivers/cell/spu/spu_main.c @@ -63,7 +63,8 @@ one_time_init(void) * This will normally be overriden by a code-gen'd function * unless CELL_FORCE_FRAGMENT_OPS_FALLBACK is set. */ - spu.fragment_ops = spu_fallback_fragment_ops; + spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops; + spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops; } @@ -90,7 +91,8 @@ main(main_param_t speid, main_param_t argp) ASSERT(sizeof(tile_t) == TILE_SIZE * TILE_SIZE * 4); ASSERT(sizeof(struct cell_command_render) % 8 == 0); - ASSERT(((unsigned long) &spu.fragment_ops_code) % 8 == 0); + ASSERT(((unsigned long) &spu.fragment_ops_code_front) % 8 == 0); + ASSERT(((unsigned long) &spu.fragment_ops_code_back) % 8 == 0); ASSERT(((unsigned long) &spu.fragment_program_code) % 8 == 0); one_time_init(); diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h index 692790c9f3..24cf7d77ce 100644 --- a/src/gallium/drivers/cell/spu/spu_main.h +++ b/src/gallium/drivers/cell/spu/spu_main.h @@ -85,8 +85,7 @@ typedef void (*spu_fragment_ops_func)(uint x, uint y, vector float fragGreen, vector float fragBlue, vector float fragAlpha, - vector unsigned int mask, - uint facing); + vector unsigned int mask); /** Function for running fragment program */ typedef vector unsigned int (*spu_fragment_program_func)(vector float *inputs, @@ -170,9 +169,10 @@ struct spu_global ubyte ztile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB; /** Current fragment ops machine code, at 8-byte boundary */ - uint fragment_ops_code[SPU_MAX_FRAGMENT_OPS_INSTS] ALIGN8_ATTRIB; - /** Current fragment ops function */ - spu_fragment_ops_func fragment_ops; + uint fragment_ops_code_front[SPU_MAX_FRAGMENT_OPS_INSTS] ALIGN8_ATTRIB; + uint fragment_ops_code_back[SPU_MAX_FRAGMENT_OPS_INSTS] ALIGN8_ATTRIB; + /** Current fragment ops functions, 0 = frontfacing, 1 = backfacing */ + spu_fragment_ops_func fragment_ops[2]; /** Current fragment program machine code, at 8-byte boundary */ uint fragment_program_code[SPU_MAX_FRAGMENT_PROGRAM_INSTS] ALIGN8_ATTRIB; diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c index f8ffc70492..683664e8a4 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.c +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c @@ -75,8 +75,7 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fragG, vector float fragB, vector float fragA, - vector unsigned int mask, - uint facing) + vector unsigned int mask) { vector float frag_aos[4]; unsigned int fbc0, fbc1, fbc2, fbc3 ; /* framebuffer/tile colors */ diff --git a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h index a61689c83a..f817abf046 100644 --- a/src/gallium/drivers/cell/spu/spu_per_fragment_op.h +++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h @@ -38,8 +38,7 @@ spu_fallback_fragment_ops(uint x, uint y, vector float fragGreen, vector float fragBlue, vector float fragAlpha, - vector unsigned int mask, - uint facing); + vector unsigned int mask); #endif /* SPU_PER_FRAGMENT_OP */ diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c index 5f908159bb..22e51a86ae 100644 --- a/src/gallium/drivers/cell/spu/spu_tri.c +++ b/src/gallium/drivers/cell/spu/spu_tri.c @@ -275,15 +275,20 @@ emit_quad( int x, int y, mask_t mask) /* Execute per-fragment/quad operations, including: * alpha test, z test, stencil test, blend and framebuffer writing. + * Note that there are two different fragment operations functions + * that can be called, one for front-facing fragments, and one + * for back-facing fragments. (Often the two are the same; + * but in some cases, like two-sided stenciling, they can be + * very different.) So choose the correct function depending + * on the calculated facing. */ - spu.fragment_ops(ix, iy, &spu.ctile, &spu.ztile, + spu.fragment_ops[setup.facing](ix, iy, &spu.ctile, &spu.ztile, fragZ, outputs[0*4+0], outputs[0*4+1], outputs[0*4+2], outputs[0*4+3], - mask, - setup.facing); + mask); } } } @@ -519,7 +524,14 @@ setup_sort_vertices(const struct vertex_header *v0, setup.oneOverArea = 1.0f / area; - /* The product of area * sign indicates front/back orientation (0/1) */ + /* The product of area * sign indicates front/back orientation (0/1). + * Just in case someone gets the bright idea of switching the front + * and back constants without noticing that we're assuming their + * values in this operation, also assert that the values are + * what we think they are. + */ + ASSERT(CELL_FACING_FRONT == 0); + ASSERT(CELL_FACING_BACK == 1); setup.facing = (area * sign > 0.0f) ^ (spu.rasterizer.front_winding == PIPE_WINDING_CW); -- cgit v1.2.3