Merge commit 'origin/gallium-0.1' into gallium-0.2

Conflicts:

	progs/fp/Makefile

15 files changed, 742 insertions, 29 deletions

diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
index 61010e4333..a04cc6c4ff 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
@@ -473,21 +473,48 @@ EMIT_R   (spe_mtspr, 0x10c);
 void
 spe_load_float(struct spe_function *p, unsigned rT, float x)
 {
-   union {
-      float f;
-      unsigned u;
-   } bits;
-   bits.f = x;
-   spe_ilhu(p, rT, bits.u >> 16);
-   spe_iohl(p, rT, bits.u & 0xffff);
+   if (x == 0.0f) {
+      spe_il(p, rT, 0x0);
+   }
+   else if (x == 0.5f) {
+      spe_ilhu(p, rT, 0x3f00);
+   }
+   else if (x == 1.0f) {
+      spe_ilhu(p, rT, 0x3f80);
+   }
+   else if (x == -1.0f) {
+      spe_ilhu(p, rT, 0xbf80);
+   }
+   else {
+      union {
+         float f;
+         unsigned u;
+      } bits;
+      bits.f = x;
+      spe_ilhu(p, rT, bits.u >> 16);
+      spe_iohl(p, rT, bits.u & 0xffff);
+   }
 }
 
 
 void
 spe_load_int(struct spe_function *p, unsigned rT, int i)
 {
-   spe_ilhu(p, rT, i >> 16);
-   spe_iohl(p, rT, i & 0xffff);
+   if (-32768 <= i && i <= 32767) {
+      spe_il(p, rT, i);
+   }
+   else {
+      spe_ilhu(p, rT, i >> 16);
+      spe_iohl(p, rT, i & 0xffff);
+   }
+}
+
+
+void
+spe_splat(struct spe_function *p, unsigned rT, unsigned rA)
+{
+   spe_ila(p, rT, 66051);
+   spe_shufb(p, rT, rA, rA, rT);
 }
 
 
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
index dee8c55c4a..d95e5aace3 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
@@ -292,6 +292,10 @@ spe_load_float(struct spe_function *p, unsigned rT, float x);
 extern void
 spe_load_int(struct spe_function *p, unsigned rT, int i);
 
+/** Replicate word 0 of rA across rT. */
+extern void
+spe_splat(struct spe_function *p, unsigned rT, unsigned rA);
+
 /** Complement/invert all bits in rT. */
 extern void
 spe_complement(struct spe_function *p, unsigned rT);
diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index e989d8c2e5..cb0631baf5 100644
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -92,6 +92,7 @@
 #define CELL_CMD_STATE_UNIFORMS      16
 #define CELL_CMD_STATE_VS_ARRAY_INFO 17
 #define CELL_CMD_STATE_BIND_VS       18
+#define CELL_CMD_STATE_FRAGMENT_PROGRAM 19
 #define CELL_CMD_STATE_ATTRIB_FETCH  20
 #define CELL_CMD_VS_EXECUTE          22
 #define CELL_CMD_FLUSH_BUFFER_RANGE  23
@@ -125,6 +126,20 @@ struct cell_command_fragment_ops
 };
 
 
+/** Max instructions for fragment programs */
+#define SPU_MAX_FRAGMENT_PROGRAM_INSTS 128
+
+/**
+ * Command to send a fragment progra to SPUs.
+ */
+struct cell_command_fragment_program
+{
+   uint64_t opcode;      /**< CELL_CMD_STATE_FRAGMENT_PROGRAM */
+   uint num_inst;        /**< Number of instructions */
+   unsigned code[SPU_MAX_FRAGMENT_PROGRAM_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 8699f3f8ec..b28f4c5c31 100644
--- a/src/gallium/drivers/cell/ppu/Makefile
+++ b/src/gallium/drivers/cell/ppu/Makefile
@@ -26,6 +26,7 @@ SOURCES = \
 	cell_draw_arrays.c \
 	cell_flush.c \
 	cell_gen_fragment.c \
+	cell_gen_fp.c \
 	cell_state_derived.c \
 	cell_state_emit.c \
 	cell_state_shader.c \
diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h
index 8cec9f45b2..14914b9c6f 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.h
+++ b/src/gallium/drivers/cell/ppu/cell_context.h
@@ -61,6 +61,7 @@ struct cell_fragment_shader_state
 {
    struct pipe_shader_state shader;
    struct tgsi_shader_info info;
+   struct spe_function code;
    void *data;
 };
 
diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c
new file mode 100644
index 0000000000..6ffe94eb14
--- /dev/null
+++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c
@@ -0,0 +1,523 @@
+/**************************************************************************
+ * 
+ * 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 fragment program/shader code.
+ *
+ * Note that we generate SOA-style code here.  So each TGSI instruction
+ * operates on four pixels (and is translated into four SPU instructions,
+ * generally speaking).
+ *
+ * \author Brian Paul
+ */
+
+
+#include "pipe/p_defines.h"
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "tgsi/tgsi_parse.h"
+#include "tgsi/tgsi_util.h"
+#include "tgsi/tgsi_exec.h"
+#include "tgsi/tgsi_dump.h"
+#include "rtasm/rtasm_ppc_spe.h"
+#include "util/u_memory.h"
+#include "cell_context.h"
+#include "cell_gen_fp.h"
+
+
+/** Set to 1 to enable debug/disassembly printfs */
+#define DISASSEM 01
+
+
+/**
+ * Context needed during code generation.
+ */
+struct codegen
+{
+   int inputs_reg;      /**< 1st function parameter */
+   int outputs_reg;     /**< 2nd function parameter */
+   int constants_reg;   /**< 3rd function parameter */
+   int temp_regs[8][4]; /**< maps TGSI temps to SPE registers */
+
+   int one_reg;         /**< register containing {1.0, 1.0, 1.0, 1.0} */
+
+   /** Per-instruction temps / intermediate temps */
+   int num_itemps;
+   int itemps[3];
+
+   struct spe_function *f;
+   boolean error;
+};
+
+
+/**
+ * Allocate an intermediate temporary register.
+ */
+static int
+get_itemp(struct codegen *gen)
+{
+   int t = spe_allocate_available_register(gen->f);
+   assert(gen->num_itemps < Elements(gen->itemps));
+   gen->itemps[gen->num_itemps++] = t;
+   return t;
+}
+
+/**
+ * Free all intermediate temporary registers.  To be called after each
+ * instruction has been emitted.
+ */
+static void
+free_itemps(struct codegen *gen)
+{
+   int i;
+   for (i = 0; i < gen->num_itemps; i++) {
+      spe_release_register(gen->f, gen->itemps[i]);
+   }
+   gen->num_itemps = 0;
+}
+
+
+/**
+ * Return index of an SPE register containing {1.0, 1.0, 1.0, 1.0}.
+ * The register is allocated and initialized upon the first call.
+ */
+static int
+get_const_one_reg(struct codegen *gen)
+{
+   if (gen->one_reg <= 0) {
+      gen->one_reg = spe_allocate_available_register(gen->f);
+   }
+
+   /* one = {1.0, 1.0, 1.0, 1.0} */
+   spe_load_float(gen->f, gen->one_reg, 1.0f);
+#if DISASSEM
+   printf("il\tr%d, 1.0f\n", gen->one_reg);
+#endif
+
+   return gen->one_reg;
+}
+
+
+/**
+ * Return the index of the SPU temporary containing the named TGSI
+ * source register.  If the TGSI register is a TGSI_FILE_TEMPORARY we
+ * just return the corresponding SPE register.  If the TGIS register
+ * is TGSI_FILE_INPUT/CONSTANT/IMMEDIATE we allocate a new SPE register
+ * and emit an SPE load instruction.
+ */
+static int
+get_src_reg(struct codegen *gen,
+            int channel,
+            const struct tgsi_full_src_register *src)
+{
+   int reg;
+
+   /* XXX need to examine src swizzle info here.
+    * That will involve changing the channel var...
+    */
+
+
+   switch (src->SrcRegister.File) {
+   case TGSI_FILE_TEMPORARY:
+      reg = gen->temp_regs[src->SrcRegister.Index][channel];
+      break;
+   case TGSI_FILE_INPUT:
+      {
+         /* offset is measured in quadwords, not bytes */
+         int offset = src->SrcRegister.Index * 4 + channel;
+         reg = get_itemp(gen);
+         /* Load:  reg = memory[(machine_reg) + offset] */
+         spe_lqd(gen->f, reg, gen->inputs_reg, offset);
+#if DISASSEM
+         printf("lqd\tr%d, r%d + %d\n", reg, gen->inputs_reg, offset);
+#endif
+      }
+      break;
+   case TGSI_FILE_IMMEDIATE:
+      /* xxx fall-through for now / fix */
+   case TGSI_FILE_CONSTANT:
+      /* xxx fall-through for now / fix */
+   default:
+      assert(0);
+   }
+
+   return reg;
+}
+
+
+/**
+ * Return the index of an SPE register to use for the given TGSI register.
+ * If the TGSI register is TGSI_FILE_TEMPORARAY, the index of the
+ * corresponding SPE register is returned.  If the TGSI register is
+ * TGSI_FILE_OUTPUT we allocate an intermediate temporary register.
+ * See store_dest_reg() below...
+ */
+static int
+get_dst_reg(struct codegen *gen,
+            int channel,
+            const struct tgsi_full_dst_register *dest)
+{
+   int reg;
+
+   switch (dest->DstRegister.File) {
+   case TGSI_FILE_TEMPORARY:
+      reg = gen->temp_regs[dest->DstRegister.Index][channel];
+      break;
+   case TGSI_FILE_OUTPUT:
+      reg = get_itemp(gen);
+      break;
+   default:
+      assert(0);
+   }
+
+   return reg;
+}
+
+
+/**
+ * When a TGSI instruction is writing to an output register, this
+ * function emits the SPE store instruction to store the value_reg.
+ * \param value_reg  the SPE register containing the value to store.
+ *                   This would have been returned by get_dst_reg().
+ */
+static void
+store_dest_reg(struct codegen *gen,
+               int value_reg, int channel,
+               const struct tgsi_full_dst_register *dest)
+{
+   switch (dest->DstRegister.File) {
+   case TGSI_FILE_TEMPORARY:
+      /* no-op */
+      break;
+   case TGSI_FILE_OUTPUT:
+      {
+         /* offset is measured in quadwords, not bytes */
+         int offset = dest->DstRegister.Index * 4 + channel;
+         /* Store: memory[(machine_reg) + offset] = reg */
+         spe_stqd(gen->f, value_reg, gen->outputs_reg, offset);
+#if DISASSEM
+         printf("stqd\tr%d, r%d + %d\n", value_reg, gen->outputs_reg, offset);
+#endif
+      }
+      break;
+   default:
+      assert(0);
+   }
+}
+
+
+static boolean
+emit_MOV(struct codegen *gen, const struct tgsi_full_instruction *inst)
+{
+   int ch;
+   for (ch = 0; ch < 4; ch++) {
+      if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) {
+         int src_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]);
+         int dst_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]);
+         /* XXX we don't always need to actually emit a mov instruction here */
+         spe_move(gen->f, dst_reg, src_reg);
+#if DISASSEM
+         printf("mov\tr%d, r%d\n", dst_reg, src_reg);
+#endif
+         store_dest_reg(gen, dst_reg, ch, &inst->FullDstRegisters[0]);
+         free_itemps(gen);
+      }
+   }
+   return true;
+}
+
+
+/**
+ * Emit addition instructions.  Recall that a single TGSI_OPCODE_ADD
+ * becomes (up to) four SPU "fa" instructions because we're doing SOA
+ * processing.
+ */
+static boolean
+emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst)
+{
+   int ch;
+   /* Loop over Red/Green/Blue/Alpha channels */
+   for (ch = 0; ch < 4; ch++) {
+      /* If the dest R, G, B or A writemask is enabled... */
+      if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) {
+         /* get indexes of the two src, one dest SPE registers */
+         int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]);
+         int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]);
+         int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]);
+
+         /* Emit actual SPE instruction: d = s1 + s2 */
+         spe_fa(gen->f, d_reg, s1_reg, s2_reg);
+#if DISASSEM
+         printf("fa\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg);
+#endif
+
+         /* Store the result (a no-op for TGSI_FILE_TEMPORARY dests) */
+         store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]);
+         /* Free any intermediate temps we allocated */
+         free_itemps(gen);
+      }
+   }
+   return true;
+}
+
+
+/**
+ * Emit multiply.  See emit_ADD for comments.
+ */
+static boolean
+emit_MUL(struct codegen *gen, const struct tgsi_full_instruction *inst)
+{
+   int ch;
+   for (ch = 0; ch < 4; ch++) {
+      if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) {
+         int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]);
+         int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]);
+         int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]);
+         /* d = s1 * s2 */
+         spe_fm(gen->f, d_reg, s1_reg, s2_reg);
+#if DISASSEM
+         printf("fm\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg);
+#endif
+         store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]);
+         free_itemps(gen);
+      }
+   }
+   return true;
+}
+
+
+/**
+ * Emit set-if-greater-than.
+ * Note that the SPE fcgt instruction produces 0x0 and 0xffffffff as
+ * the result but OpenGL/TGSI needs 0.0 and 1.0 results.
+ * We can easily convert 0x0/0xffffffff to 0.0/1.0 with a bitwise AND.
+ */
+static boolean
+emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst)
+{
+   int ch;
+
+   for (ch = 0; ch < 4; ch++) {
+      if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) {
+         int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]);
+         int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]);
+         int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]);
+
+         /* d = (s1 > s2) */
+         spe_fcgt(gen->f, d_reg, s1_reg, s2_reg);
+#if DISASSEM
+         printf("fcgt\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg);
+#endif
+
+         /* convert d from 0x0/0xffffffff to 0.0/1.0 */
+         /* d = d & one_reg */
+         spe_and(gen->f, d_reg, d_reg, get_const_one_reg(gen));
+#if DISASSEM
+         printf("and\tr%d, r%d, r%d\n", d_reg, d_reg, get_const_one_reg(gen));
+#endif
+
+         store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]);
+         free_itemps(gen);
+      }
+   }
+
+   return true;
+}
+
+
+/**
+ * Emit END instruction.
+ * We just return from the shader function at this point.
+ *
+ * Note that there may be more code after this that would be
+ * called by TGSI_OPCODE_CALL.
+ */
+static boolean
+emit_END(struct codegen *gen)
+{
+   /* return from function call */
+   spe_bi(gen->f, SPE_REG_RA, 0, 0);
+#if DISASSEM
+   printf("bi\trRA\n");
+#endif
+   return true;
+}
+
+
+/**
+ * Emit code for the given instruction.  Just a big switch stmt.
+ */
+static boolean
+emit_instruction(struct codegen *gen,
+                 const struct tgsi_full_instruction *inst)
+{
+   switch (inst->Instruction.Opcode) {
+   case TGSI_OPCODE_MOV:
+      return emit_MOV(gen, inst);
+   case TGSI_OPCODE_MUL:
+      return emit_MUL(gen, inst);
+   case TGSI_OPCODE_ADD:
+      return emit_ADD(gen, inst);
+   case TGSI_OPCODE_SGT:
+      return emit_SGT(gen, inst);
+   case TGSI_OPCODE_END:
+      return emit_END(gen);
+
+   /* XXX lots more cases to do... */
+
+   default:
+      return false;
+   }
+
+   return true;
+}
+
+
+
+/**
+ * Emit "code" for a TGSI declaration.
+ * We only care about TGSI TEMPORARY register declarations at this time.
+ * For each TGSI TEMPORARY we allocate four SPE registers.
+ */
+static void
+emit_declaration(struct codegen *gen, const struct tgsi_full_declaration *decl)
+{
+   int i, ch;
+
+   switch (decl->Declaration.File) {
+   case TGSI_FILE_TEMPORARY:
+#if DISASSEM
+      printf("Declare temp reg %d .. %d\n",
+             decl->DeclarationRange.First,
+             decl->DeclarationRange.Last);
+#endif
+      for (i = decl->DeclarationRange.First;
+           i <= decl->DeclarationRange.Last;
+           i++) {
+         for (ch = 0; ch < 4; ch++) {
+            gen->temp_regs[i][ch] = spe_allocate_available_register(gen->f);
+         }
+
+         /* XXX if we run out of SPE registers, we need to spill
+          * to SPU memory.  someday...
+          */
+
+#if DISASSEM
+         printf("  SPE regs: %d %d %d %d\n",
+                gen->temp_regs[i][0],
+                gen->temp_regs[i][1],
+                gen->temp_regs[i][2],
+                gen->temp_regs[i][3]);
+#endif
+      }
+      break;
+   default:
+      ; /* ignore */
+   }
+}
+
+
+/**
+ * Translate TGSI shader code to SPE instructions.  This is done when
+ * the state tracker gives us a new shader (via pipe->create_fs_state()).
+ *
+ * \param cell    the rendering context (in)
+ * \param tokens  the TGSI shader (in)
+ * \param f       the generated function (out)
+ */
+boolean
+cell_gen_fragment_program(struct cell_context *cell,
+                          const struct tgsi_token *tokens,
+                          struct spe_function *f)
+{
+   struct tgsi_parse_context parse;
+   struct codegen gen;
+
+   memset(&gen, 0, sizeof(gen));
+   gen.f = f;
+
+   /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */
+   gen.inputs_reg = 3;     /* pointer to inputs array */
+   gen.outputs_reg = 4;    /* pointer to outputs array */
+   gen.constants_reg = 5;  /* pointer to constants array */
+
+   spe_init_func(f, SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE);
+   spe_allocate_register(f, gen.inputs_reg);
+   spe_allocate_register(f, gen.outputs_reg);
+   spe_allocate_register(f, gen.constants_reg);
+
+#if DISASSEM
+   printf("Begin %s\n", __FUNCTION__);
+   tgsi_dump(tokens, 0);
+#endif
+
+   tgsi_parse_init(&parse, tokens);
+
+   while (!tgsi_parse_end_of_tokens(&parse) && !gen.error) {
+      tgsi_parse_token(&parse);
+
+      switch (parse.FullToken.Token.Type) {
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+#if 0
+         if (!note_immediate(&gen, &parse.FullToken.FullImmediate ))
+            goto fail;
+#endif
+         break;
+
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         emit_declaration(&gen, &parse.FullToken.FullDeclaration);
+         break;
+
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         if (!emit_instruction(&gen, &parse.FullToken.FullInstruction )) {
+            gen.error = true;
+         }
+         break;
+
+      default:
+         assert(0);
+
+      }
+   }
+
+
+   if (gen.error) {
+      /* terminate the SPE code */
+      return emit_END(&gen);
+   }
+
+#if DISASSEM
+   printf("cell_gen_fragment_program nr instructions: %d\n", f->num_inst);
+   printf("End %s\n", __FUNCTION__);
+#endif
+
+   tgsi_parse_free( &parse );
+
+   return !gen.error;
+}
diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.h b/src/gallium/drivers/cell/ppu/cell_gen_fp.h
new file mode 100644
index 0000000000..99faea7046
--- /dev/null
+++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.h
@@ -0,0 +1,42 @@
+/**************************************************************************
+ * 
+ * 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_FP_H
+#define CELL_GEN_FP_H
+
+
+
+extern boolean
+cell_gen_fragment_program(struct cell_context *cell,
+                          const struct tgsi_token *tokens,
+                          struct spe_function *f);
+
+
+#endif /* CELL_GEN_FP_H */
+
diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c
index 79a82ef72b..06219d4e98 100644
--- a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c
+++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c
@@ -265,6 +265,8 @@ gen_blend(const struct pipe_blend_state *blend,
    int one_reg = spe_allocate_available_register(f);
    int tmp_reg = spe_allocate_available_register(f);
 
+   boolean one_reg_set = false; /* avoid setting one_reg more than once */
+
    ASSERT(blend->blend_enable);
 
    /* Unpack/convert framebuffer colors from four 32-bit packed colors
@@ -275,7 +277,7 @@ gen_blend(const struct pipe_blend_state *blend,
       int mask_reg = spe_allocate_available_register(f);
 
       /* mask = {0x000000ff, 0x000000ff, 0x000000ff, 0x000000ff} */
-      spe_fsmbi(f, mask_reg, 0x1111);
+      spe_load_int(f, mask_reg, 0xff);
 
       /* XXX there may be more clever ways to implement the following code */
       switch (color_format) {
@@ -418,7 +420,10 @@ gen_blend(const struct pipe_blend_state *blend,
       break;
    case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
       /* one = {1.0, 1.0, 1.0, 1.0} */
-      spe_load_float(f, one_reg, 1.0f);
+      if (!one_reg_set) {
+         spe_load_float(f, one_reg, 1.0f);
+         one_reg_set = true;
+      }
       /* tmp = one - fragA */
       spe_fs(f, tmp_reg, one_reg, fragA_reg);
       /* term = fb * tmp */
@@ -446,7 +451,10 @@ gen_blend(const struct pipe_blend_state *blend,
       break;
    case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
       /* one = {1.0, 1.0, 1.0, 1.0} */
-      spe_load_float(f, one_reg, 1.0f);
+      if (!one_reg_set) {
+         spe_load_float(f, one_reg, 1.0f);
+         one_reg_set = true;
+      }
       /* tmp = one - fragA */
       spe_fs(f, tmp_reg, one_reg, fragA_reg);
       /* termA = fbA * tmp */
@@ -616,7 +624,7 @@ gen_pack_colors(struct spe_function *f,
  * \param f     the generated function (out)
  */
 void
-gen_fragment_function(struct cell_context *cell, struct spe_function *f)
+cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f)
 {
    const struct pipe_depth_stencil_alpha_state *dsa =
       &cell->depth_stencil->base;
@@ -850,7 +858,7 @@ gen_fragment_function(struct cell_context *cell, struct spe_function *f)
       spe_release_register(f, rgba_reg);
    }
 
-   printf("gen_fragment_ops nr instructions: %u\n", f->num_inst);
+   //printf("gen_fragment_ops nr instructions: %u\n", f->num_inst);
 
    spe_bi(f, SPE_REG_RA, 0, 0);  /* return from function call */
 
diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h
index 0ea0fc690c..b59de198dc 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
-gen_fragment_function(struct cell_context *cell, struct spe_function *f);
+cell_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 180b89c1f6..2da3097983 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_emit.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c
@@ -73,6 +73,22 @@ cell_emit_state(struct cell_context *cell)
 #endif
    }
 
+   if (cell->dirty & (CELL_NEW_FS)) {
+      /* Send new fragment program to SPUs */
+      struct cell_command_fragment_program *fp
+            = cell_batch_alloc(cell, sizeof(*fp));
+      fp->opcode = CELL_CMD_STATE_FRAGMENT_PROGRAM;
+      fp->num_inst = cell->fs->code.num_inst;
+      memcpy(&fp->code, cell->fs->code.store,
+             SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE);
+      if (0) {
+         int i;
+         printf("PPU Emit CELL_CMD_STATE_FRAGMENT_PROGRAM:\n");
+         for (i = 0; i < fp->num_inst; i++) {
+            printf(" %3d: 0x%08x\n", i, fp->code[i]);
+         }
+      }
+   }
 
    if (cell->dirty & (CELL_NEW_FRAMEBUFFER |
                       CELL_NEW_DEPTH_STENCIL |
@@ -85,7 +101,7 @@ cell_emit_state(struct cell_context *cell)
       struct spe_function spe_code;
 
       /* generate new code */
-      gen_fragment_function(cell, &spe_code);
+      cell_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,
diff --git a/src/gallium/drivers/cell/ppu/cell_state_shader.c b/src/gallium/drivers/cell/ppu/cell_state_shader.c
index 97e44eeb1a..3a0d066da2 100644
--- a/src/gallium/drivers/cell/ppu/cell_state_shader.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_shader.c
@@ -34,7 +34,7 @@
 
 #include "cell_context.h"
 #include "cell_state.h"
-
+#include "cell_gen_fp.h"
 
 
 /** cast wrapper */
@@ -61,7 +61,7 @@ static void *
 cell_create_fs_state(struct pipe_context *pipe,
                      const struct pipe_shader_state *templ)
 {
-   /*struct cell_context *cell = cell_context(pipe);*/
+   struct cell_context *cell = cell_context(pipe);
    struct cell_fragment_shader_state *cfs;
 
    cfs = CALLOC_STRUCT(cell_fragment_shader_state);
@@ -76,6 +76,8 @@ cell_create_fs_state(struct pipe_context *pipe,
 
    tgsi_scan_shader(templ->tokens, &cfs->info);
 
+   cell_gen_fragment_program(cell, cfs->shader.tokens, &cfs->code);
+
    return cfs;
 }
 
@@ -102,6 +104,8 @@ cell_delete_fs_state(struct pipe_context *pipe, void *fs)
 {
    struct cell_fragment_shader_state *cfs = cell_fragment_shader_state(fs);
 
+   spe_release_func(&cfs->code);
+
    FREE((void *) cfs->shader.tokens);
    FREE(cfs);
 }
diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c
index 2a7cb75f59..78260c4259 100644
--- a/src/gallium/drivers/cell/spu/spu_main.c
+++ b/src/gallium/drivers/cell/spu/spu_main.c
@@ -232,7 +232,7 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops)
       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);
-   /* Copy state info */
+   /* 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));
 
@@ -245,6 +245,21 @@ cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops)
 
 
 static void
+cmd_state_fragment_program(const struct cell_command_fragment_program *fp)
+{
+   if (Debug)
+      printf("SPU %u: CMD_STATE_FRAGMENT_PROGRAM\n", spu.init.id);
+   /* Copy SPU code from batch buffer to spu buffer */
+   memcpy(spu.fragment_program_code, fp->code,
+          SPU_MAX_FRAGMENT_PROGRAM_INSTS * 4);
+#if 01
+   /* Point function pointer at new code */
+   spu.fragment_program = (spu_fragment_program_func)spu.fragment_program_code;
+#endif
+}
+
+
+static void
 cmd_state_framebuffer(const struct cell_command_framebuffer *cmd)
 {
    if (Debug)
@@ -473,6 +488,14 @@ cmd_batch(uint opcode)
             pos += sizeof(*fops) / 8;
          }
          break;
+      case CELL_CMD_STATE_FRAGMENT_PROGRAM:
+         {
+            struct cell_command_fragment_program *fp
+               = (struct cell_command_fragment_program *) &buffer[pos];
+            cmd_state_fragment_program(fp);
+            pos += sizeof(*fp) / 8;
+         }
+         break;
       case CELL_CMD_STATE_SAMPLER:
          {
             struct cell_command_sampler *sampler
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
index d40539da83..2c7b625840 100644
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -75,6 +75,12 @@ typedef void (*spu_fragment_ops_func)(uint x, uint y,
                                       vector float fragAlpha,
                                       vector unsigned int mask);
 
+/** Function for running fragment program */
+typedef void (*spu_fragment_program_func)(vector float *inputs,
+                                          vector float *outputs,
+                                          vector float *constants);
+
+
 struct spu_framebuffer
 {
    void *color_start;              /**< addr of color surface in main memory */
@@ -142,9 +148,18 @@ struct spu_global
    /** Current fragment ops function */
    spu_fragment_ops_func fragment_ops;
 
+   /** Current fragment program machine code */
+   uint fragment_program_code[SPU_MAX_FRAGMENT_PROGRAM_INSTS];
+   /** Current fragment ops function */
+   spu_fragment_program_func fragment_program;
+
    /** Current texture sampler function */
    spu_sample_texture_func sample_texture[CELL_MAX_SAMPLERS];
 
+   /** Fragment program constants (XXX preliminary/used) */
+#define MAX_CONSTANTS 32
+   vector float constants[MAX_CONSTANTS];
+
 } ALIGN16_ATTRIB;
 
 
diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c
index f02cdd1f76..8b93878192 100644
--- a/src/gallium/drivers/cell/spu/spu_tri.c
+++ b/src/gallium/drivers/cell/spu/spu_tri.c
@@ -314,7 +314,42 @@ emit_quad( int x, int y, mask_t mask )
       }
       else {
          /* simple shading */
+#if 0
          eval_coeff(1, (float) x, (float) y, colors);
+
+#else
+         /* XXX new fragment program code */
+
+         if (spu.fragment_program) {
+            vector float inputs[4*4], outputs[2*4];
+
+            /* setup inputs */
+            eval_coeff(1, (float) x, (float) y, inputs);
+
+            /* Execute the current fragment program */
+            spu.fragment_program(inputs, outputs, spu.constants);
+
+            /* Copy outputs */
+            colors[0] = outputs[0*4+0];
+            colors[1] = outputs[0*4+1];
+            colors[2] = outputs[0*4+2];
+            colors[3] = outputs[0*4+3];
+
+            if (0 && spu.init.id==0 && y == 48) {
+               printf("colors[0] = %f %f %f %f\n",
+                      spu_extract(colors[0], 0),
+                      spu_extract(colors[0], 1),
+                      spu_extract(colors[0], 2),
+                      spu_extract(colors[0], 3));
+               printf("colors[1] = %f %f %f %f\n",
+                      spu_extract(colors[1], 0),
+                      spu_extract(colors[1], 1),
+                      spu_extract(colors[1], 2),
+                      spu_extract(colors[1], 3));
+            }
+
+         }
+#endif
       }
 
 
diff --git a/src/gallium/winsys/xlib/xm_winsys.c b/src/gallium/winsys/xlib/xm_winsys.c
index c4a30d3702..2acbc94fc8 100644
--- a/src/gallium/winsys/xlib/xm_winsys.c
+++ b/src/gallium/winsys/xlib/xm_winsys.c
@@ -289,21 +289,19 @@ xm_buffer_destroy(struct pipe_winsys *pws,
  *  +--+--+
  */
 static void
-twiddle_tile(uint *tile)
+twiddle_tile(const uint *tileIn, uint *tileOut)
 {
-   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];
+         tileOut[y * TILE_SIZE + (x + 0)] = tileIn[k];
+         tileOut[y * TILE_SIZE + (x + 1)] = tileIn[k+1];
+         tileOut[(y + 1) * TILE_SIZE + (x + 0)] = tileIn[k+2];
+         tileOut[(y + 1) * TILE_SIZE + (x + 1)] = tileIn[k+3];
       }
    }
-   memcpy(tile, tile2, sizeof(tile2));
 }
 
 
@@ -339,6 +337,7 @@ xmesa_display_surface_tiled(XMesaBuffer b, const struct pipe_surface *surf)
 
    for (y = 0; y < surf->height; y += TILE_SIZE) {
       for (x = 0; x < surf->width; x += TILE_SIZE) {
+         uint tmpTile[TILE_SIZE * TILE_SIZE];
          int tx = x / TILE_SIZE;
          int ty = y / TILE_SIZE;
          int offset = ty * tilesPerRow + tx;
@@ -352,9 +351,9 @@ xmesa_display_surface_tiled(XMesaBuffer b, const struct pipe_surface *surf)
 
          offset *= 4 * TILE_SIZE * TILE_SIZE;
 
-         ximage->data = (char *) xm_buf->data + offset;
-
-         twiddle_tile((uint *) ximage->data);
+         twiddle_tile((uint *) ((char *) xm_buf->data + offset),
+                      tmpTile);
+         ximage->data = (char*) tmpTile;
 
          if (XSHM_ENABLED(xm_buf)) {
 #if defined(USE_XSHM) && !defined(XFree86Server)