cell: add support for fragment shader constant buffers

diff --git a/src/gallium/drivers/cell/common.h b/src/gallium/drivers/cell/common.h
index c223bc174493986562676a4b178f2e9f7c65a170..d261c1a6408e557343cdc0a67320c0acdae4e41c 100644 (file)
--- a/src/gallium/drivers/cell/common.h
+++ b/src/gallium/drivers/cell/common.h
@@ -94,6 +94,7 @@
 #define CELL_CMD_STATE_BIND_VS       18
 #define CELL_CMD_STATE_FRAGMENT_PROGRAM 19
 #define CELL_CMD_STATE_ATTRIB_FETCH  20
+#define CELL_CMD_STATE_FS_CONSTANTS  21
 #define CELL_CMD_VS_EXECUTE          22
 #define CELL_CMD_FLUSH_BUFFER_RANGE  23
 

diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c
index 131a2356fe5a48e3728b58a94955c8aaef3d67e3..3065869d0436b5050e511b5862681e31ee2e791e 100644 (file)
--- a/src/gallium/drivers/cell/ppu/cell_gen_fp.c
+++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c
@@ -215,7 +215,15 @@ get_src_reg(struct codegen *gen,
          reg = gen->imm_regs[src->SrcRegister.Index][swizzle];
          break;
       case TGSI_FILE_CONSTANT:
-         /* xxx fall-through for now / fix */
+         {
+            /* offset is measured in quadwords, not bytes */
+            int offset = src->SrcRegister.Index * 4 + swizzle;
+            reg = get_itemp(gen);
+            reg_is_itemp = TRUE;
+            /* Load:  reg = memory[(machine_reg) + offset] */
+            spe_lqd(gen->f, reg, gen->constants_reg, offset);
+         }
+         break;
       default:
          assert(0);
       }

diff --git a/src/gallium/drivers/cell/ppu/cell_state.h b/src/gallium/drivers/cell/ppu/cell_state.h
index a7771a55a316fbcb6676faa04a9f6186a15ab4fe..b193170f9ce975404dec7ae9271b8a3ba9b3bf54 100644 (file)
--- a/src/gallium/drivers/cell/ppu/cell_state.h
+++ b/src/gallium/drivers/cell/ppu/cell_state.h
@@ -44,8 +44,9 @@
 #define CELL_NEW_TEXTURE       0x800
 #define CELL_NEW_VERTEX        0x1000
 #define CELL_NEW_VS            0x2000
-#define CELL_NEW_CONSTANTS     0x4000
-#define CELL_NEW_VERTEX_INFO   0x8000
+#define CELL_NEW_VS_CONSTANTS  0x4000
+#define CELL_NEW_FS_CONSTANTS  0x8000
+#define CELL_NEW_VERTEX_INFO   0x10000
 
 
 extern void

diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c
index a36fd3a6015a3d788470f9908e62eca4d53472dc..cbfa393cfbae1d59ebf2099bb13a90310f73266c 100644 (file)
--- a/src/gallium/drivers/cell/ppu/cell_state_emit.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c
@@ -25,6 +25,7 @@
  * 
  **************************************************************************/
 
+#include "pipe/p_inlines.h"
 #include "util/u_memory.h"
 #include "cell_context.h"
 #include "cell_gen_fragment.h"
@@ -162,6 +163,24 @@ cell_emit_state(struct cell_context *cell)
       }
    }
 
+   if (cell->dirty & (CELL_NEW_FS_CONSTANTS)) {
+      const uint shader = PIPE_SHADER_FRAGMENT;
+      const uint num_const = cell->constants[shader].size / sizeof(float);
+      uint i, j;
+      float *buf = cell_batch_alloc(cell, 16 + num_const * sizeof(float));
+      uint64_t *ibuf = (uint64_t *) buf;
+      const float *constants = pipe_buffer_map(cell->pipe.screen,
+                                               cell->constants[shader].buffer,
+                                               PIPE_BUFFER_USAGE_CPU_READ);
+      ibuf[0] = CELL_CMD_STATE_FS_CONSTANTS;
+      ibuf[1] = num_const;
+      j = 4;
+      for (i = 0; i < num_const; i++) {
+         buf[j++] = constants[i];
+      }
+      pipe_buffer_unmap(cell->pipe.screen, cell->constants[shader].buffer);
+   }
+
    if (cell->dirty & (CELL_NEW_FRAMEBUFFER |
                       CELL_NEW_DEPTH_STENCIL |
                       CELL_NEW_BLEND)) {

diff --git a/src/gallium/drivers/cell/ppu/cell_state_shader.c b/src/gallium/drivers/cell/ppu/cell_state_shader.c
index 3a0d066da2a7591e970675bf8310c8b04df80ec7..54a17eaf2b7692cc3592a596bc619270452c3b40 100644 (file)
--- a/src/gallium/drivers/cell/ppu/cell_state_shader.c
+++ b/src/gallium/drivers/cell/ppu/cell_state_shader.c
@@ -197,7 +197,10 @@ cell_set_constant_buffer(struct pipe_context *pipe,
                         buf->buffer);
    cell->constants[shader].size = buf->size;
 
-   cell->dirty |= CELL_NEW_CONSTANTS;
+   if (shader == PIPE_SHADER_VERTEX)
+      cell->dirty |= CELL_NEW_VS_CONSTANTS;
+   else if (shader == PIPE_SHADER_FRAGMENT)
+      cell->dirty |= CELL_NEW_FS_CONSTANTS;
 }
 
 

diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index ec9da5d88705c996c67dbe92a0f9df85bb045162..91a4c137e7c4045b533bf4c37bd9f632df38731e 100644 (file)
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -231,6 +231,25 @@ cmd_state_fragment_program(const struct cell_command_fragment_program *fp)
 }
 
 
+static uint
+cmd_state_fs_constants(const uint64_t *buffer, uint pos)
+{
+   const uint num_const = buffer[pos + 1];
+   const float *constants = (const float *) &buffer[pos + 2];
+   uint i;
+
+   DEBUG_PRINTF("CMD_STATE_FS_CONSTANTS (%u)\n", num_const);
+
+   /* Expand each float to float[4] for SOA execution */
+   for (i = 0; i < num_const; i++) {
+      spu.constants[i] = spu_splats(constants[i]);
+   }
+
+   /* return new buffer pos (in 8-byte words) */
+   return pos + 2 + num_const / 2;
+}
+
+
 static void
 cmd_state_framebuffer(const struct cell_command_framebuffer *cmd)
 {
@@ -456,6 +475,9 @@ cmd_batch(uint opcode)
             pos += sizeof(*fp) / 8;
          }
          break;
+      case CELL_CMD_STATE_FS_CONSTANTS:
+         pos = cmd_state_fs_constants(buffer, pos);
+         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 1cd577c23caae9cffbdea0f2e3bd62f6cc6749a3..82c9c69a3a84871ad673a53abb88cb779791110a 100644 (file)
--- a/src/gallium/drivers/cell/spu/spu_main.h
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -41,6 +41,9 @@
 #define MAX_HEIGHT 1024
 
 
+#define CELL_MAX_CONSTANTS 32  /**< number of float[4] constants */
+
+
 /**
  * A tile is basically a TILE_SIZE x TILE_SIZE block of 4-byte pixels.
  * The data may be addressed through several different types.
@@ -157,9 +160,8 @@ struct spu_global
    /** 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];
+   /** Fragment program constants */
+   vector float constants[4 * CELL_MAX_CONSTANTS];
 
 } ALIGN16_ATTRIB;