diff options
Diffstat (limited to 'src/gallium/drivers/r300/r300_emit.c')
| -rw-r--r-- | src/gallium/drivers/r300/r300_emit.c | 583 |
1 files changed, 235 insertions, 348 deletions
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c index c329c0f30f..014b382edf 100644 --- a/src/gallium/drivers/r300/r300_emit.c +++ b/src/gallium/drivers/r300/r300_emit.c @@ -43,21 +43,11 @@ void r300_emit_blend_state(struct r300_context* r300, (struct pipe_framebuffer_state*)r300->fb_state.state; CS_LOCALS(r300); - BEGIN_CS(size); - OUT_CS_REG(R300_RB3D_ROPCNTL, blend->rop); - OUT_CS_REG_SEQ(R300_RB3D_CBLEND, 3); if (fb->nr_cbufs) { - OUT_CS(blend->blend_control); - OUT_CS(blend->alpha_blend_control); - OUT_CS(blend->color_channel_mask); + WRITE_CS_TABLE(blend->cb, size); } else { - OUT_CS(0); - OUT_CS(0); - OUT_CS(0); - /* XXX also disable fastfill here once it's supported */ + WRITE_CS_TABLE(blend->cb_no_readwrite, size); } - OUT_CS_REG(R300_RB3D_DITHER_CTL, blend->dither); - END_CS; } void r300_emit_blend_color_state(struct r300_context* r300, @@ -66,40 +56,16 @@ void r300_emit_blend_color_state(struct r300_context* r300, struct r300_blend_color_state* bc = (struct r300_blend_color_state*)state; CS_LOCALS(r300); - if (r300->screen->caps.is_r500) { - BEGIN_CS(size); - OUT_CS_REG_SEQ(R500_RB3D_CONSTANT_COLOR_AR, 2); - OUT_CS(bc->blend_color_red_alpha); - OUT_CS(bc->blend_color_green_blue); - END_CS; - } else { - BEGIN_CS(size); - OUT_CS_REG(R300_RB3D_BLEND_COLOR, bc->blend_color); - END_CS; - } + WRITE_CS_TABLE(bc->cb, size); } void r300_emit_clip_state(struct r300_context* r300, unsigned size, void* state) { - struct pipe_clip_state* clip = (struct pipe_clip_state*)state; + struct r300_clip_state* clip = (struct r300_clip_state*)state; CS_LOCALS(r300); - if (r300->screen->caps.has_tcl) { - BEGIN_CS(size); - OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG, - (r300->screen->caps.is_r500 ? - R500_PVS_UCP_START : R300_PVS_UCP_START)); - OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, 6 * 4); - OUT_CS_TABLE(clip->ucp, 6 * 4); - OUT_CS_REG(R300_VAP_CLIP_CNTL, ((1 << clip->nr) - 1) | - R300_PS_UCP_MODE_CLIP_AS_TRIFAN); - END_CS; - } else { - BEGIN_CS(size); - OUT_CS_REG(R300_VAP_CLIP_CNTL, R300_CLIP_DISABLE); - END_CS; - } + WRITE_CS_TABLE(clip->cb, size); } void r300_emit_dsa_state(struct r300_context* r300, unsigned size, void* state) @@ -107,27 +73,13 @@ void r300_emit_dsa_state(struct r300_context* r300, unsigned size, void* state) struct r300_dsa_state* dsa = (struct r300_dsa_state*)state; struct pipe_framebuffer_state* fb = (struct pipe_framebuffer_state*)r300->fb_state.state; - struct pipe_stencil_ref stencil_ref = r300->stencil_ref; CS_LOCALS(r300); - BEGIN_CS(size); - OUT_CS_REG(R300_FG_ALPHA_FUNC, dsa->alpha_function); - OUT_CS_REG_SEQ(R300_ZB_CNTL, 3); - if (fb->zsbuf) { - OUT_CS(dsa->z_buffer_control); - OUT_CS(dsa->z_stencil_control); + WRITE_CS_TABLE(&dsa->cb_begin, size); } else { - OUT_CS(0); - OUT_CS(0); + WRITE_CS_TABLE(dsa->cb_no_readwrite, size); } - - OUT_CS(dsa->stencil_ref_mask | stencil_ref.ref_value[0]); - - if (r300->screen->caps.is_r500) { - OUT_CS_REG(R500_ZB_STENCILREFMASK_BF, dsa->stencil_ref_bf | stencil_ref.ref_value[1]); - } - END_CS; } static const float * get_rc_constant_state( @@ -175,7 +127,7 @@ static const float * get_rc_constant_state( /* Convert a normal single-precision float into the 7.16 format * used by the R300 fragment shader. */ -static uint32_t pack_float24(float f) +uint32_t pack_float24(float f) { union { float fl; @@ -206,101 +158,27 @@ static uint32_t pack_float24(float f) return float24; } -unsigned r300_get_fs_atom_size(struct r300_context *r300) -{ - struct r300_fragment_shader *fs = r300_fs(r300); - unsigned imm_count = fs->shader->immediates_count; - struct r300_fragment_program_code *code = &fs->shader->code.code.r300; - - return 19 + - code->alu.length * 4 + - (code->tex.length ? (1 + code->tex.length) : 0) + - (imm_count ? imm_count * 5 : 0); -} - void r300_emit_fs(struct r300_context* r300, unsigned size, void *state) { struct r300_fragment_shader *fs = r300_fs(r300); - struct rX00_fragment_program_code* generic_code = &fs->shader->code; - struct r300_fragment_program_code * code = &generic_code->code.r300; - unsigned i; - unsigned imm_count = fs->shader->immediates_count; - unsigned imm_first = fs->shader->externals_count; - unsigned imm_end = generic_code->constants.Count; - struct rc_constant *constants = generic_code->constants.Constants; CS_LOCALS(r300); - BEGIN_CS(size); - OUT_CS_REG(R300_US_CONFIG, code->config); - OUT_CS_REG(R300_US_PIXSIZE, code->pixsize); - OUT_CS_REG(R300_US_CODE_OFFSET, code->code_offset); - - OUT_CS_REG_SEQ(R300_US_CODE_ADDR_0, 4); - OUT_CS_TABLE(code->code_addr, 4); - - OUT_CS_REG_SEQ(R300_US_ALU_RGB_INST_0, code->alu.length); - for (i = 0; i < code->alu.length; i++) - OUT_CS(code->alu.inst[i].rgb_inst); - - OUT_CS_REG_SEQ(R300_US_ALU_RGB_ADDR_0, code->alu.length); - for (i = 0; i < code->alu.length; i++) - OUT_CS(code->alu.inst[i].rgb_addr); - - OUT_CS_REG_SEQ(R300_US_ALU_ALPHA_INST_0, code->alu.length); - for (i = 0; i < code->alu.length; i++) - OUT_CS(code->alu.inst[i].alpha_inst); - - OUT_CS_REG_SEQ(R300_US_ALU_ALPHA_ADDR_0, code->alu.length); - for (i = 0; i < code->alu.length; i++) - OUT_CS(code->alu.inst[i].alpha_addr); - - if (code->tex.length) { - OUT_CS_REG_SEQ(R300_US_TEX_INST_0, code->tex.length); - OUT_CS_TABLE(code->tex.inst, code->tex.length); - } - - /* Emit immediates. */ - if (imm_count) { - for(i = imm_first; i < imm_end; ++i) { - if (constants[i].Type == RC_CONSTANT_IMMEDIATE) { - const float *data = constants[i].u.Immediate; - - OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4); - OUT_CS(pack_float24(data[0])); - OUT_CS(pack_float24(data[1])); - OUT_CS(pack_float24(data[2])); - OUT_CS(pack_float24(data[3])); - } - } - } - - OUT_CS_REG(R300_FG_DEPTH_SRC, fs->shader->fg_depth_src); - OUT_CS_REG(R300_US_W_FMT, fs->shader->us_out_w); - END_CS; + WRITE_CS_TABLE(fs->shader->cb_code, fs->shader->cb_code_size); } void r300_emit_fs_constants(struct r300_context* r300, unsigned size, void *state) { struct r300_fragment_shader *fs = r300_fs(r300); - struct rc_constant_list *constants = &fs->shader->code.constants; struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state; - unsigned i, count = fs->shader->externals_count; + unsigned count = fs->shader->externals_count * 4; CS_LOCALS(r300); if (count == 0) return; BEGIN_CS(size); - OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X, count * 4); - for(i = 0; i < count; ++i) { - const float *data; - assert(constants->Constants[i].Type == RC_CONSTANT_EXTERNAL); - data = buf->constants[i]; - OUT_CS(pack_float24(data[0])); - OUT_CS(pack_float24(data[1])); - OUT_CS(pack_float24(data[2])); - OUT_CS(pack_float24(data[3])); - } + OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X, count); + OUT_CS_TABLE(buf->constants, count); END_CS; } @@ -312,6 +190,8 @@ void r300_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, vo unsigned count = fs->shader->rc_state_count; unsigned first = fs->shader->externals_count; unsigned end = constants->Count; + uint32_t cdata[4]; + unsigned j; CS_LOCALS(r300); if (count == 0) @@ -323,85 +203,29 @@ void r300_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, vo const float *data = get_rc_constant_state(r300, &constants->Constants[i]); + for (j = 0; j < 4; j++) + cdata[j] = pack_float24(data[j]); + OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4); - OUT_CS(pack_float24(data[0])); - OUT_CS(pack_float24(data[1])); - OUT_CS(pack_float24(data[2])); - OUT_CS(pack_float24(data[3])); + OUT_CS_TABLE(cdata, 4); } } END_CS; } -unsigned r500_get_fs_atom_size(struct r300_context *r300) -{ - struct r300_fragment_shader *fs = r300_fs(r300); - unsigned imm_count = fs->shader->immediates_count; - struct r500_fragment_program_code *code = &fs->shader->code.code.r500; - - return 17 + - ((code->inst_end + 1) * 6) + - (imm_count ? imm_count * 7 : 0); -} - void r500_emit_fs(struct r300_context* r300, unsigned size, void *state) { struct r300_fragment_shader *fs = r300_fs(r300); - struct rX00_fragment_program_code* generic_code = &fs->shader->code; - struct r500_fragment_program_code * code = &generic_code->code.r500; - unsigned i; - unsigned imm_count = fs->shader->immediates_count; - unsigned imm_first = fs->shader->externals_count; - unsigned imm_end = generic_code->constants.Count; - struct rc_constant *constants = generic_code->constants.Constants; CS_LOCALS(r300); - BEGIN_CS(size); - OUT_CS_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO); - OUT_CS_REG(R500_US_PIXSIZE, code->max_temp_idx); - OUT_CS_REG(R500_US_CODE_RANGE, - R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(code->inst_end)); - OUT_CS_REG(R500_US_CODE_OFFSET, 0); - OUT_CS_REG(R500_US_CODE_ADDR, - R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(code->inst_end)); - - OUT_CS_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_INSTR); - OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, (code->inst_end + 1) * 6); - for (i = 0; i <= code->inst_end; i++) { - OUT_CS(code->inst[i].inst0); - OUT_CS(code->inst[i].inst1); - OUT_CS(code->inst[i].inst2); - OUT_CS(code->inst[i].inst3); - OUT_CS(code->inst[i].inst4); - OUT_CS(code->inst[i].inst5); - } - - /* Emit immediates. */ - if (imm_count) { - for(i = imm_first; i < imm_end; ++i) { - if (constants[i].Type == RC_CONSTANT_IMMEDIATE) { - const float *data = constants[i].u.Immediate; - - OUT_CS_REG(R500_GA_US_VECTOR_INDEX, - R500_GA_US_VECTOR_INDEX_TYPE_CONST | - (i & R500_GA_US_VECTOR_INDEX_MASK)); - OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, 4); - OUT_CS_TABLE(data, 4); - } - } - } - - OUT_CS_REG(R300_FG_DEPTH_SRC, fs->shader->fg_depth_src); - OUT_CS_REG(R300_US_W_FMT, fs->shader->us_out_w); - END_CS; + WRITE_CS_TABLE(fs->shader->cb_code, fs->shader->cb_code_size); } void r500_emit_fs_constants(struct r300_context* r300, unsigned size, void *state) { struct r300_fragment_shader *fs = r300_fs(r300); - struct rc_constant_list *constants = &fs->shader->code.constants; struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state; - unsigned i, count = fs->shader->externals_count; + unsigned count = fs->shader->externals_count * 4; CS_LOCALS(r300); if (count == 0) @@ -409,11 +233,8 @@ void r500_emit_fs_constants(struct r300_context* r300, unsigned size, void *stat BEGIN_CS(size); OUT_CS_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_CONST); - OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, count * 4); - for(i = 0; i < count; ++i) { - assert(constants->Constants[i].Type == RC_CONSTANT_EXTERNAL); - } - OUT_CS_TABLE(buf->constants, count * 4); + OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, count); + OUT_CS_TABLE(buf->constants, count); END_CS; } @@ -446,83 +267,170 @@ void r500_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, vo END_CS; } +void r300_emit_gpu_flush(struct r300_context *r300, unsigned size, void *state) +{ + struct r300_gpu_flush *gpuflush = (struct r300_gpu_flush*)state; + struct pipe_framebuffer_state* fb = + (struct pipe_framebuffer_state*)r300->fb_state.state; + CS_LOCALS(r300); + + BEGIN_CS(size); + + /* Set up scissors. + * By writing to the SC registers, SC & US assert idle. */ + OUT_CS_REG_SEQ(R300_SC_SCISSORS_TL, 2); + if (r300->screen->caps.is_r500) { + OUT_CS(0); + OUT_CS(((fb->width - 1) << R300_SCISSORS_X_SHIFT) | + ((fb->height - 1) << R300_SCISSORS_Y_SHIFT)); + } else { + OUT_CS((1440 << R300_SCISSORS_X_SHIFT) | + (1440 << R300_SCISSORS_Y_SHIFT)); + OUT_CS(((fb->width + 1440-1) << R300_SCISSORS_X_SHIFT) | + ((fb->height + 1440-1) << R300_SCISSORS_Y_SHIFT)); + } + + /* Flush CB & ZB caches and wait until the 3D engine is idle and clean. */ + OUT_CS_TABLE(gpuflush->cb_flush_clean, 6); + END_CS; +} + +void r300_emit_aa_state(struct r300_context *r300, unsigned size, void *state) +{ + struct r300_aa_state *aa = (struct r300_aa_state*)state; + CS_LOCALS(r300); + + BEGIN_CS(size); + OUT_CS_REG(R300_GB_AA_CONFIG, aa->aa_config); + + if (aa->dest) { + OUT_CS_REG_SEQ(R300_RB3D_AARESOLVE_OFFSET, 1); + OUT_CS_RELOC(aa->dest->buffer, aa->dest->offset, 0, aa->dest->domain, 0); + + OUT_CS_REG_SEQ(R300_RB3D_AARESOLVE_PITCH, 1); + OUT_CS_RELOC(aa->dest->buffer, aa->dest->pitch, 0, aa->dest->domain, 0); + } + + OUT_CS_REG(R300_RB3D_AARESOLVE_CTL, aa->aaresolve_ctl); + END_CS; +} + void r300_emit_fb_state(struct r300_context* r300, unsigned size, void* state) { struct pipe_framebuffer_state* fb = (struct pipe_framebuffer_state*)state; - struct r300_texture* tex; - struct pipe_surface* surf; - int i; + struct r300_surface* surf; + unsigned i; CS_LOCALS(r300); BEGIN_CS(size); - /* Flush and free renderbuffer caches. */ - OUT_CS_REG(R300_RB3D_DSTCACHE_CTLSTAT, - R300_RB3D_DSTCACHE_CTLSTAT_DC_FREE_FREE_3D_TAGS | - R300_RB3D_DSTCACHE_CTLSTAT_DC_FLUSH_FLUSH_DIRTY_3D); - OUT_CS_REG(R300_ZB_ZCACHE_CTLSTAT, - R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE | - R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE); - - /* Set the number of colorbuffers. */ - if (fb->nr_cbufs > 1) { - if (r300->screen->caps.is_r500) { - OUT_CS_REG(R300_RB3D_CCTL, - R300_RB3D_CCTL_NUM_MULTIWRITES(fb->nr_cbufs) | - R300_RB3D_CCTL_INDEPENDENT_COLORFORMAT_ENABLE_ENABLE); - } else { - OUT_CS_REG(R300_RB3D_CCTL, - R300_RB3D_CCTL_NUM_MULTIWRITES(fb->nr_cbufs)); - } + /* NUM_MULTIWRITES replicates COLOR[0] to all colorbuffers, which is not + * what we usually want. */ + if (r300->screen->caps.is_r500) { + OUT_CS_REG(R300_RB3D_CCTL, + R300_RB3D_CCTL_INDEPENDENT_COLORFORMAT_ENABLE_ENABLE); } else { - OUT_CS_REG(R300_RB3D_CCTL, 0x0); + OUT_CS_REG(R300_RB3D_CCTL, 0); } /* Set up colorbuffers. */ for (i = 0; i < fb->nr_cbufs; i++) { - surf = fb->cbufs[i]; - tex = r300_texture(surf->texture); - assert(tex && tex->buffer && "cbuf is marked, but NULL!"); + surf = r300_surface(fb->cbufs[i]); OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1); - OUT_CS_TEX_RELOC(tex, surf->offset, 0, tex->domain, 0); + OUT_CS_RELOC(surf->buffer, surf->offset, 0, surf->domain, 0); OUT_CS_REG_SEQ(R300_RB3D_COLORPITCH0 + (4 * i), 1); - OUT_CS_TEX_RELOC(tex, tex->fb_state.colorpitch[surf->level], - 0, tex->domain, 0); - - OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), tex->fb_state.us_out_fmt); - } - for (; i < 4; i++) { - OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), R300_US_OUT_FMT_UNUSED); + OUT_CS_RELOC(surf->buffer, surf->pitch, 0, surf->domain, 0); } /* Set up a zbuffer. */ if (fb->zsbuf) { - surf = fb->zsbuf; - tex = r300_texture(surf->texture); - assert(tex && tex->buffer && "zsbuf is marked, but NULL!"); + surf = r300_surface(fb->zsbuf); - OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1); - OUT_CS_TEX_RELOC(tex, surf->offset, 0, tex->domain, 0); + OUT_CS_REG(R300_ZB_FORMAT, surf->format); - OUT_CS_REG(R300_ZB_FORMAT, tex->fb_state.zb_format); + OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1); + OUT_CS_RELOC(surf->buffer, surf->offset, 0, surf->domain, 0); OUT_CS_REG_SEQ(R300_ZB_DEPTHPITCH, 1); - OUT_CS_TEX_RELOC(tex, tex->fb_state.depthpitch[surf->level], - 0, tex->domain, 0); + OUT_CS_RELOC(surf->buffer, surf->pitch, 0, surf->domain, 0); + + /* HiZ RAM. */ + if (r300->screen->caps.has_hiz) { + OUT_CS_REG(R300_ZB_HIZ_OFFSET, 0); + OUT_CS_REG(R300_ZB_HIZ_PITCH, 0); + } + + /* Z Mask RAM. (compressed zbuffer) */ + OUT_CS_REG(R300_ZB_ZMASK_OFFSET, 0); + OUT_CS_REG(R300_ZB_ZMASK_PITCH, 0); } - OUT_CS_REG_SEQ(R300_SC_SCISSORS_TL, 2); - if (r300->screen->caps.is_r500) { - OUT_CS(0); - OUT_CS(((fb->width - 1) << R300_SCISSORS_X_SHIFT) | - ((fb->height - 1) << R300_SCISSORS_Y_SHIFT)); - } else { - OUT_CS((1440 << R300_SCISSORS_X_SHIFT) | - (1440 << R300_SCISSORS_Y_SHIFT)); - OUT_CS(((fb->width + 1440-1) << R300_SCISSORS_X_SHIFT) | - ((fb->height + 1440-1) << R300_SCISSORS_Y_SHIFT)); + END_CS; +} + +void r300_emit_hyperz_state(struct r300_context *r300, + unsigned size, void *state) +{ + CS_LOCALS(r300); + WRITE_CS_TABLE(state, size); +} + +void r300_emit_fb_state_pipelined(struct r300_context *r300, + unsigned size, void *state) +{ + struct pipe_framebuffer_state* fb = + (struct pipe_framebuffer_state*)r300->fb_state.state; + unsigned i; + CS_LOCALS(r300); + + BEGIN_CS(size); + + /* Colorbuffer format in the US block. + * (must be written after unpipelined regs) */ + OUT_CS_REG_SEQ(R300_US_OUT_FMT_0, 4); + for (i = 0; i < fb->nr_cbufs; i++) { + OUT_CS(r300_surface(fb->cbufs[i])->format); + } + for (; i < 4; i++) { + OUT_CS(R300_US_OUT_FMT_UNUSED); + } + + /* Multisampling. Depends on framebuffer sample count. + * These are pipelined regs and as such cannot be moved + * to the AA state. */ + if (r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0)) { + unsigned mspos0 = 0x66666666; + unsigned mspos1 = 0x6666666; + + if (fb->nr_cbufs && fb->cbufs[0]->texture->nr_samples > 1) { + /* Subsample placement. These may not be optimal. */ + switch (fb->cbufs[0]->texture->nr_samples) { + case 2: + mspos0 = 0x33996633; + mspos1 = 0x6666663; + break; + case 3: + mspos0 = 0x33936933; + mspos1 = 0x6666663; + break; + case 4: + mspos0 = 0x33939933; + mspos1 = 0x3966663; + break; + case 6: + mspos0 = 0x22a2aa22; + mspos1 = 0x2a65672; + break; + default: + debug_printf("r300: Bad number of multisamples!\n"); + } + } + + OUT_CS_REG_SEQ(R300_GB_MSPOS0, 2); + OUT_CS(mspos0); + OUT_CS(mspos1); } END_CS; } @@ -544,13 +452,14 @@ void r300_emit_query_start(struct r300_context *r300, unsigned size, void*state) OUT_CS_REG(R300_ZB_ZPASS_DATA, 0); END_CS; query->begin_emitted = TRUE; + query->flushed = FALSE; } - static void r300_emit_query_end_frag_pipes(struct r300_context *r300, struct r300_query *query) { struct r300_capabilities* caps = &r300->screen->caps; + struct r300_winsys_buffer *buf = r300->query_current->buffer; CS_LOCALS(r300); assert(caps->num_frag_pipes); @@ -569,28 +478,28 @@ static void r300_emit_query_end_frag_pipes(struct r300_context *r300, /* pipe 3 only */ OUT_CS_REG(R300_SU_REG_DEST, 1 << 3); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 3), - 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 3) * 4, + 0, query->domain, 0); case 3: /* pipe 2 only */ OUT_CS_REG(R300_SU_REG_DEST, 1 << 2); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 2), - 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 2) * 4, + 0, query->domain, 0); case 2: /* pipe 1 only */ /* As mentioned above, accomodate RV380 and older. */ OUT_CS_REG(R300_SU_REG_DEST, 1 << (caps->high_second_pipe ? 3 : 1)); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 1), - 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 1) * 4, + 0, query->domain, 0); case 1: /* pipe 0 only */ OUT_CS_REG(R300_SU_REG_DEST, 1 << 0); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 0), - 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 0) * 4, + 0, query->domain, 0); break; default: fprintf(stderr, "r300: Implementation error: Chipset reports %d" @@ -606,12 +515,13 @@ static void r300_emit_query_end_frag_pipes(struct r300_context *r300, static void rv530_emit_query_end_single_z(struct r300_context *r300, struct r300_query *query) { + struct r300_winsys_buffer *buf = r300->query_current->buffer; CS_LOCALS(r300); BEGIN_CS(8); OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, query->num_results * 4, 0, query->domain, 0); OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL); END_CS; } @@ -619,15 +529,16 @@ static void rv530_emit_query_end_single_z(struct r300_context *r300, static void rv530_emit_query_end_double_z(struct r300_context *r300, struct r300_query *query) { + struct r300_winsys_buffer *buf = r300->query_current->buffer; CS_LOCALS(r300); BEGIN_CS(14); OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 0) * 4, 0, query->domain, 0); OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_1); OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1); - OUT_CS_BUF_RELOC(r300->oqbo, query->offset + sizeof(uint32_t), 0, r300_buffer(r300->oqbo)->domain, 0); + OUT_CS_RELOC(buf, (query->num_results + 1) * 4, 0, query->domain, 0); OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL); END_CS; } @@ -652,57 +563,36 @@ void r300_emit_query_end(struct r300_context* r300) r300_emit_query_end_frag_pipes(r300, query); query->begin_emitted = FALSE; + query->num_results += query->num_pipes; + + /* XXX grab all the results and reset the counter. */ + if (query->num_results >= query->buffer_size / 4 - 4) { + query->num_results = (query->buffer_size / 4) / 2; + fprintf(stderr, "r300: Rewinding OQBO...\n"); + } +} + +void r300_emit_invariant_state(struct r300_context *r300, + unsigned size, void *state) +{ + CS_LOCALS(r300); + WRITE_CS_TABLE(state, size); } void r300_emit_rs_state(struct r300_context* r300, unsigned size, void* state) { - struct r300_rs_state* rs = (struct r300_rs_state*)state; - float scale, offset; + struct r300_rs_state* rs = state; CS_LOCALS(r300); BEGIN_CS(size); - OUT_CS_REG(R300_VAP_CNTL_STATUS, rs->vap_control_status); - - OUT_CS_REG(R300_GB_AA_CONFIG, rs->antialiasing_config); - - OUT_CS_REG(R300_GA_POINT_SIZE, rs->point_size); - OUT_CS_REG_SEQ(R300_GA_POINT_MINMAX, 2); - OUT_CS(rs->point_minmax); - OUT_CS(rs->line_control); - + OUT_CS_TABLE(rs->cb_main, 25); if (rs->polygon_offset_enable) { - scale = rs->depth_scale * 12; - offset = rs->depth_offset; - - switch (r300->zbuffer_bpp) { - case 16: - offset *= 4; - break; - case 24: - offset *= 2; - break; + if (r300->zbuffer_bpp == 16) { + OUT_CS_TABLE(rs->cb_poly_offset_zb16, 5); + } else { + OUT_CS_TABLE(rs->cb_poly_offset_zb24, 5); } - - OUT_CS_REG_SEQ(R300_SU_POLY_OFFSET_FRONT_SCALE, 4); - OUT_CS_32F(scale); - OUT_CS_32F(offset); - OUT_CS_32F(scale); - OUT_CS_32F(offset); } - - OUT_CS_REG_SEQ(R300_SU_POLY_OFFSET_ENABLE, 2); - OUT_CS(rs->polygon_offset_enable); - OUT_CS(rs->cull_mode); - OUT_CS_REG(R300_GA_LINE_STIPPLE_CONFIG, rs->line_stipple_config); - OUT_CS_REG(R300_GA_LINE_STIPPLE_VALUE, rs->line_stipple_value); - OUT_CS_REG(R300_GA_POLY_MODE, rs->polygon_mode); - OUT_CS_REG(R300_SC_CLIP_RULE, rs->clip_rule); - OUT_CS_REG(R300_GB_ENABLE, rs->stuffing_enable); - OUT_CS_REG_SEQ(R300_GA_POINT_S0, 4); - OUT_CS_32F(rs->point_texcoord_left); - OUT_CS_32F(rs->point_texcoord_bottom); - OUT_CS_32F(rs->point_texcoord_right); - OUT_CS_32F(rs->point_texcoord_top); END_CS; } @@ -717,9 +607,18 @@ void r300_emit_rs_block_state(struct r300_context* r300, if (SCREEN_DBG_ON(r300->screen, DBG_DRAW)) { r500_dump_rs_block(rs); - } - DBG(r300, DBG_DRAW, "r300: RS emit:\n"); + fprintf(stderr, "r300: RS emit:\n"); + + for (i = 0; i < count; i++) + fprintf(stderr, " : ip %d: 0x%08x\n", i, rs->ip[i]); + + for (i = 0; i < count; i++) + fprintf(stderr, " : inst %d: 0x%08x\n", i, rs->inst[i]); + + fprintf(stderr, " : count: 0x%08x inst_count: 0x%08x\n", + rs->count, rs->inst_count); + } BEGIN_CS(size); OUT_CS_REG_SEQ(R300_VAP_VTX_STATE_CNTL, 2); @@ -735,9 +634,6 @@ void r300_emit_rs_block_state(struct r300_context* r300, OUT_CS_REG_SEQ(R300_RS_IP_0, count); } OUT_CS_TABLE(rs->ip, count); - for (i = 0; i < count; i++) { - DBG(r300, DBG_DRAW, " : ip %d: 0x%08x\n", i, rs->ip[i]); - } OUT_CS_REG_SEQ(R300_RS_COUNT, 2); OUT_CS(rs->count); @@ -749,13 +645,6 @@ void r300_emit_rs_block_state(struct r300_context* r300, OUT_CS_REG_SEQ(R300_RS_INST_0, count); } OUT_CS_TABLE(rs->inst, count); - for (i = 0; i < count; i++) { - DBG(r300, DBG_DRAW, " : inst %d: 0x%08x\n", i, rs->inst[i]); - } - - DBG(r300, DBG_DRAW, " : count: 0x%08x inst_count: 0x%08x\n", - rs->count, rs->inst_count); - END_CS; } @@ -821,19 +710,11 @@ void r300_emit_aos(struct r300_context* r300, int offset, boolean indexed) struct pipe_vertex_element *velem = r300->velems->velem; struct r300_buffer *buf; int i; + unsigned *hw_format_size = r300->velems->hw_format_size; unsigned size1, size2, aos_count = r300->velems->count; unsigned packet_size = (aos_count * 3 + 1) / 2; CS_LOCALS(r300); - for (i = 0; i < aos_count; i++) { - if ((vbuf[velem[i].vertex_buffer_index].buffer_offset + velem[i].src_offset) % 4 != 0) { - /* XXX We must align the buffer. */ - assert(0); - fprintf(stderr, "r300: Unaligned vertex buffer offsets aren't supported, aborting..\n"); - abort(); - } - } - BEGIN_CS(2 + packet_size + aos_count * 2); OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, packet_size); OUT_CS(aos_count | (!indexed ? R300_VC_FORCE_PREFETCH : 0)); @@ -841,8 +722,8 @@ void r300_emit_aos(struct r300_context* r300, int offset, boolean indexed) for (i = 0; i < aos_count - 1; i += 2) { vb1 = &vbuf[velem[i].vertex_buffer_index]; vb2 = &vbuf[velem[i+1].vertex_buffer_index]; - size1 = util_format_get_blocksize(velem[i].src_format); - size2 = util_format_get_blocksize(velem[i+1].src_format); + size1 = hw_format_size[i]; + size2 = hw_format_size[i+1]; OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride) | R300_VBPNTR_SIZE1(size2) | R300_VBPNTR_STRIDE1(vb2->stride)); @@ -852,7 +733,7 @@ void r300_emit_aos(struct r300_context* r300, int offset, boolean indexed) if (aos_count & 1) { vb1 = &vbuf[velem[i].vertex_buffer_index]; - size1 = util_format_get_blocksize(velem[i].src_format); + size1 = hw_format_size[i]; OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride)); OUT_CS(vb1->buffer_offset + velem[i].src_offset + offset * vb1->stride); @@ -897,21 +778,25 @@ void r300_emit_vertex_stream_state(struct r300_context* r300, unsigned i; CS_LOCALS(r300); - DBG(r300, DBG_DRAW, "r300: PSC emit:\n"); + if (DBG_ON(r300, DBG_DRAW)) { + fprintf(stderr, "r300: PSC emit:\n"); + + for (i = 0; i < streams->count; i++) { + fprintf(stderr, " : prog_stream_cntl%d: 0x%08x\n", i, + streams->vap_prog_stream_cntl[i]); + } + + for (i = 0; i < streams->count; i++) { + fprintf(stderr, " : prog_stream_cntl_ext%d: 0x%08x\n", i, + streams->vap_prog_stream_cntl_ext[i]); + } + } BEGIN_CS(size); OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_0, streams->count); OUT_CS_TABLE(streams->vap_prog_stream_cntl, streams->count); - for (i = 0; i < streams->count; i++) { - DBG(r300, DBG_DRAW, " : prog_stream_cntl%d: 0x%08x\n", i, - streams->vap_prog_stream_cntl[i]); - } OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_EXT_0, streams->count); OUT_CS_TABLE(streams->vap_prog_stream_cntl_ext, streams->count); - for (i = 0; i < streams->count; i++) { - DBG(r300, DBG_DRAW, " : prog_stream_cntl_ext%d: 0x%08x\n", i, - streams->vap_prog_stream_cntl_ext[i]); - } END_CS; } @@ -924,6 +809,13 @@ void r300_emit_pvs_flush(struct r300_context* r300, unsigned size, void* state) END_CS; } +void r300_emit_vap_invariant_state(struct r300_context *r300, + unsigned size, void *state) +{ + CS_LOCALS(r300); + WRITE_CS_TABLE(state, size); +} + void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state) { struct r300_vertex_shader* vs = (struct r300_vertex_shader*)state; @@ -948,6 +840,7 @@ void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state) CS_LOCALS(r300); BEGIN_CS(size); + /* R300_VAP_PVS_CODE_CNTL_0 * R300_VAP_PVS_CONST_CNTL * R300_VAP_PVS_CODE_CNTL_1 @@ -1012,12 +905,7 @@ void r300_emit_viewport_state(struct r300_context* r300, BEGIN_CS(size); OUT_CS_REG_SEQ(R300_SE_VPORT_XSCALE, 6); - OUT_CS_32F(viewport->xscale); - OUT_CS_32F(viewport->xoffset); - OUT_CS_32F(viewport->yscale); - OUT_CS_32F(viewport->yoffset); - OUT_CS_32F(viewport->zscale); - OUT_CS_32F(viewport->zoffset); + OUT_CS_TABLE(&viewport->xscale, 6); OUT_CS_REG(R300_VAP_VTE_CNTL, viewport->vte_control); END_CS; } @@ -1099,10 +987,9 @@ validate: } } /* ...occlusion query buffer... */ - if (r300->query_start.dirty || - (r300->query_current && r300->query_current->begin_emitted)) { - if (!r300_add_buffer(r300->rws, r300->oqbo, - 0, r300_buffer(r300->oqbo)->domain)) { + if (r300->query_current) { + if (!r300->rws->add_buffer(r300->rws, r300->query_current->buffer, + 0, r300->query_current->domain)) { r300->context.flush(&r300->context, 0, NULL); goto validate; } |