aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium/drivers/cell/spu
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/drivers/cell/spu')
-rw-r--r--src/gallium/drivers/cell/spu/.gitignore1
-rw-r--r--src/gallium/drivers/cell/spu/Makefile82
-rw-r--r--src/gallium/drivers/cell/spu/spu_colorpack.h145
-rw-r--r--src/gallium/drivers/cell/spu/spu_command.c815
-rw-r--r--src/gallium/drivers/cell/spu/spu_command.h35
-rw-r--r--src/gallium/drivers/cell/spu/spu_dcache.c145
-rw-r--r--src/gallium/drivers/cell/spu/spu_dcache.h37
-rw-r--r--src/gallium/drivers/cell/spu/spu_exec.c1933
-rw-r--r--src/gallium/drivers/cell/spu/spu_exec.h172
-rw-r--r--src/gallium/drivers/cell/spu/spu_funcs.c173
-rw-r--r--src/gallium/drivers/cell/spu/spu_funcs.h35
-rw-r--r--src/gallium/drivers/cell/spu/spu_main.c117
-rw-r--r--src/gallium/drivers/cell/spu/spu_main.h254
-rw-r--r--src/gallium/drivers/cell/spu/spu_per_fragment_op.c631
-rw-r--r--src/gallium/drivers/cell/spu/spu_per_fragment_op.h44
-rw-r--r--src/gallium/drivers/cell/spu/spu_render.c295
-rw-r--r--src/gallium/drivers/cell/spu/spu_render.h38
-rw-r--r--src/gallium/drivers/cell/spu/spu_shuffle.h186
-rw-r--r--src/gallium/drivers/cell/spu/spu_texture.c641
-rw-r--r--src/gallium/drivers/cell/spu/spu_texture.h67
-rw-r--r--src/gallium/drivers/cell/spu/spu_tile.c126
-rw-r--r--src/gallium/drivers/cell/spu/spu_tile.h75
-rw-r--r--src/gallium/drivers/cell/spu/spu_tri.c809
-rw-r--r--src/gallium/drivers/cell/spu/spu_tri.h37
-rw-r--r--src/gallium/drivers/cell/spu/spu_util.c167
-rw-r--r--src/gallium/drivers/cell/spu/spu_vertex_fetch.c145
-rw-r--r--src/gallium/drivers/cell/spu/spu_vertex_shader.c244
-rw-r--r--src/gallium/drivers/cell/spu/spu_vertex_shader.h66
28 files changed, 7515 insertions, 0 deletions
diff --git a/src/gallium/drivers/cell/spu/.gitignore b/src/gallium/drivers/cell/spu/.gitignore
new file mode 100644
index 00000000000..2be9a2d3242
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/.gitignore
@@ -0,0 +1 @@
+g3d_spu
diff --git a/src/gallium/drivers/cell/spu/Makefile b/src/gallium/drivers/cell/spu/Makefile
new file mode 100644
index 00000000000..116453b79c5
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/Makefile
@@ -0,0 +1,82 @@
+# Gallium3D Cell driver: SPU code
+
+# This makefile builds the g3d_spu.a file that's linked into the
+# PPU code/library.
+
+
+TOP = ../../../../..
+include $(TOP)/configs/current
+
+
+PROG = g3d
+
+PROG_SPU = $(PROG)_spu
+PROG_SPU_A = $(PROG)_spu.a
+PROG_SPU_EMBED_O = $(PROG)_spu-embed.o
+
+
+SOURCES = \
+ spu_command.c \
+ spu_dcache.c \
+ spu_funcs.c \
+ spu_main.c \
+ spu_per_fragment_op.c \
+ spu_render.c \
+ spu_texture.c \
+ spu_tile.c \
+ spu_tri.c
+
+OLD_SOURCES = \
+ spu_exec.c \
+ spu_util.c \
+ spu_vertex_fetch.c \
+ spu_vertex_shader.c
+
+
+SPU_OBJECTS = $(SOURCES:.c=.o) \
+
+SPU_ASM_OUT = $(SOURCES:.c=.s) \
+
+INCLUDE_DIRS = \
+ -I$(TOP)/src/mesa \
+ -I$(TOP)/src/gallium/include \
+ -I$(TOP)/src/gallium/auxiliary \
+ -I$(TOP)/src/gallium/drivers
+
+
+.c.o:
+ $(SPU_CC) $(SPU_CFLAGS) -c $<
+
+.c.s:
+ $(SPU_CC) $(SPU_CFLAGS) -O3 -S $<
+
+
+# The .a file will be linked into the main/PPU executable
+default: $(PROG_SPU_A)
+
+$(PROG_SPU_A): $(PROG_SPU_EMBED_O)
+ $(SPU_AR) $(SPU_AR_FLAGS) $(PROG_SPU_A) $(PROG_SPU_EMBED_O)
+
+$(PROG_SPU_EMBED_O): $(PROG_SPU)
+ $(SPU_EMBED) $(SPU_EMBED_FLAGS) $(PROG_SPU) $(PROG_SPU) $(PROG_SPU_EMBED_O)
+
+$(PROG_SPU): $(SPU_OBJECTS)
+ $(SPU_CC) -o $(PROG_SPU) $(SPU_OBJECTS) $(SPU_LFLAGS)
+
+
+
+asmfiles: $(SPU_ASM_OUT)
+
+
+clean:
+ rm -f *~ *.o *.a *.d *.s $(PROG_SPU)
+
+
+
+depend: $(SOURCES)
+ rm -f depend
+ touch depend
+ $(MKDEP) $(MKDEP_OPTIONS) $(INCLUDE_DIRS) $(SOURCES) 2> /dev/null
+
+include depend
+
diff --git a/src/gallium/drivers/cell/spu/spu_colorpack.h b/src/gallium/drivers/cell/spu/spu_colorpack.h
new file mode 100644
index 00000000000..d7ce0055248
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_colorpack.h
@@ -0,0 +1,145 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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 SPU_COLORPACK_H
+#define SPU_COLORPACK_H
+
+
+#include <transpose_matrix4x4.h>
+#include <spu_intrinsics.h>
+
+
+static INLINE unsigned int
+spu_pack_R8G8B8A8(vector float rgba)
+{
+ vector unsigned int out = spu_convtu(rgba, 32);
+
+ out = spu_shuffle(out, out, ((vector unsigned char) {
+ 0, 4, 8, 12, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0 }) );
+
+ return spu_extract(out, 0);
+}
+
+
+static INLINE unsigned int
+spu_pack_A8R8G8B8(vector float rgba)
+{
+ vector unsigned int out = spu_convtu(rgba, 32);
+ out = spu_shuffle(out, out, ((vector unsigned char) {
+ 12, 0, 4, 8, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0}) );
+ return spu_extract(out, 0);
+}
+
+
+static INLINE unsigned int
+spu_pack_B8G8R8A8(vector float rgba)
+{
+ vector unsigned int out = spu_convtu(rgba, 32);
+ out = spu_shuffle(out, out, ((vector unsigned char) {
+ 8, 4, 0, 12, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0}) );
+ return spu_extract(out, 0);
+}
+
+
+static INLINE unsigned int
+spu_pack_color_shuffle(vector float rgba, vector unsigned char shuffle)
+{
+ vector unsigned int out = spu_convtu(rgba, 32);
+ out = spu_shuffle(out, out, shuffle);
+ return spu_extract(out, 0);
+}
+
+
+static INLINE vector float
+spu_unpack_B8G8R8A8(uint color)
+{
+ vector unsigned int color_u4 = spu_splats(color);
+ color_u4 = spu_shuffle(color_u4, color_u4,
+ ((vector unsigned char) {
+ 2, 2, 2, 2,
+ 1, 1, 1, 1,
+ 0, 0, 0, 0,
+ 3, 3, 3, 3}) );
+ return spu_convtf(color_u4, 32);
+}
+
+
+static INLINE vector float
+spu_unpack_A8R8G8B8(uint color)
+{
+ vector unsigned int color_u4 = spu_splats(color);
+ color_u4 = spu_shuffle(color_u4, color_u4,
+ ((vector unsigned char) {
+ 1, 1, 1, 1,
+ 2, 2, 2, 2,
+ 3, 3, 3, 3,
+ 0, 0, 0, 0}) );
+ return spu_convtf(color_u4, 32);
+}
+
+
+/**
+ * \param color_in - array of 32-bit packed ARGB colors
+ * \param color_out - returns float colors in RRRR, GGGG, BBBB, AAAA order
+ */
+static INLINE void
+spu_unpack_A8R8G8B8_transpose4(const vector unsigned int color_in[4],
+ vector float color_out[4])
+{
+ vector unsigned int c0;
+
+ c0 = spu_shuffle(color_in[0], color_in[0],
+ ((vector unsigned char) {
+ 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) );
+ color_out[0] = spu_convtf(c0, 32);
+
+ c0 = spu_shuffle(color_in[1], color_in[1],
+ ((vector unsigned char) {
+ 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) );
+ color_out[1] = spu_convtf(c0, 32);
+
+ c0 = spu_shuffle(color_in[2], color_in[2],
+ ((vector unsigned char) {
+ 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) );
+ color_out[2] = spu_convtf(c0, 32);
+
+ c0 = spu_shuffle(color_in[3], color_in[3],
+ ((vector unsigned char) {
+ 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 0, 0}) );
+ color_out[3] = spu_convtf(c0, 32);
+
+ _transpose_matrix4x4(color_out, color_out);
+}
+
+
+
+#endif /* SPU_COLORPACK_H */
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
new file mode 100644
index 00000000000..5c0179d9546
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -0,0 +1,815 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+
+/**
+ * SPU command processing code
+ */
+
+
+#include <stdio.h>
+#include <libmisc.h>
+
+#include "pipe/p_defines.h"
+
+#include "spu_command.h"
+#include "spu_main.h"
+#include "spu_render.h"
+#include "spu_per_fragment_op.h"
+#include "spu_texture.h"
+#include "spu_tile.h"
+#include "spu_vertex_shader.h"
+#include "spu_dcache.h"
+#include "cell/common.h"
+
+
+struct spu_vs_context draw;
+
+
+/**
+ * Buffers containing dynamically generated SPU code:
+ */
+static unsigned char attribute_fetch_code_buffer[136 * PIPE_MAX_ATTRIBS]
+ ALIGN16_ATTRIB;
+
+
+
+static INLINE int
+align(int value, int alignment)
+{
+ return (value + alignment - 1) & ~(alignment - 1);
+}
+
+
+
+/**
+ * Tell the PPU that this SPU has finished copying a buffer to
+ * local store and that it may be reused by the PPU.
+ * This is done by writting a 16-byte batch-buffer-status block back into
+ * main memory (in cell_context->buffer_status[]).
+ */
+static void
+release_buffer(uint buffer)
+{
+ /* Evidently, using less than a 16-byte status doesn't work reliably */
+ static const vector unsigned int status = {CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE,
+ CELL_BUFFER_STATUS_FREE};
+ const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer);
+ uint *dst = spu.init.buffer_status + index;
+
+ ASSERT(buffer < CELL_NUM_BUFFERS);
+
+ mfc_put((void *) &status, /* src in local memory */
+ (unsigned int) dst, /* dst in main memory */
+ sizeof(status), /* size */
+ TAG_MISC, /* tag is unimportant */
+ 0, /* tid */
+ 0 /* rid */);
+}
+
+
+/**
+ * Write CELL_FENCE_SIGNALLED back to the fence status qword in main memory.
+ * There's a qword of status per SPU.
+ */
+static void
+cmd_fence(struct cell_command_fence *fence_cmd)
+{
+ static const vector unsigned int status = {CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED,
+ CELL_FENCE_SIGNALLED};
+ uint *dst = (uint *) fence_cmd->fence;
+ dst += 4 * spu.init.id; /* main store/memory address, not local store */
+ ASSERT_ALIGN16(dst);
+ mfc_put((void *) &status, /* src in local memory */
+ (unsigned int) dst, /* dst in main memory */
+ sizeof(status), /* size */
+ TAG_FENCE, /* tag */
+ 0, /* tid */
+ 0 /* rid */);
+}
+
+
+static void
+cmd_clear_surface(const struct cell_command_clear_surface *clear)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value);
+
+ if (clear->surface == 0) {
+ spu.fb.color_clear_value = clear->value;
+ if (spu.init.debug_flags & CELL_DEBUG_CHECKER) {
+ uint x = (spu.init.id << 4) | (spu.init.id << 12) |
+ (spu.init.id << 20) | (spu.init.id << 28);
+ spu.fb.color_clear_value ^= x;
+ }
+ }
+ else {
+ spu.fb.depth_clear_value = clear->value;
+ }
+
+#define CLEAR_OPT 1
+#if CLEAR_OPT
+
+ /* Simply set all tiles' status to CLEAR.
+ * When we actually begin rendering into a tile, we'll initialize it to
+ * the clear value. If any tiles go untouched during the frame,
+ * really_clear_tiles() will set them to the clear value.
+ */
+ if (clear->surface == 0) {
+ memset(spu.ctile_status, TILE_STATUS_CLEAR, sizeof(spu.ctile_status));
+ }
+ else {
+ memset(spu.ztile_status, TILE_STATUS_CLEAR, sizeof(spu.ztile_status));
+ }
+
+#else
+
+ /*
+ * This path clears the whole framebuffer to the clear color right now.
+ */
+
+ /*
+ printf("SPU: %s num=%d w=%d h=%d\n",
+ __FUNCTION__, num_tiles, spu.fb.width_tiles, spu.fb.height_tiles);
+ */
+
+ /* init a single tile to the clear value */
+ if (clear->surface == 0) {
+ clear_c_tile(&spu.ctile);
+ }
+ else {
+ clear_z_tile(&spu.ztile);
+ }
+
+ /* walk over my tiles, writing the 'clear' tile's data */
+ {
+ const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles;
+ uint i;
+ for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) {
+ uint tx = i % spu.fb.width_tiles;
+ uint ty = i / spu.fb.width_tiles;
+ if (clear->surface == 0)
+ put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0);
+ else
+ put_tile(tx, ty, &spu.ztile, TAG_SURFACE_CLEAR, 1);
+ }
+ }
+
+ if (spu.init.debug_flags & CELL_DEBUG_SYNC) {
+ wait_on_mask(1 << TAG_SURFACE_CLEAR);
+ }
+
+#endif /* CLEAR_OPT */
+
+ D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF done\n");
+}
+
+
+static void
+cmd_release_verts(const struct cell_command_release_verts *release)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "RELEASE VERTS %u\n", release->vertex_buf);
+ ASSERT(release->vertex_buf != ~0U);
+ release_buffer(release->vertex_buf);
+}
+
+
+/**
+ * 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)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n");
+
+ /* Copy state info (for fallback case only - this will eventually
+ * go away when the fallback case goes away)
+ */
+ memcpy(&spu.depth_stencil_alpha, &fops->dsa, sizeof(fops->dsa));
+ memcpy(&spu.blend, &fops->blend, sizeof(fops->blend));
+ memcpy(&spu.blend_color, &fops->blend_color, sizeof(fops->blend_color));
+
+ /* Make sure the SPU knows which buffers it's expected to read when
+ * it's told to pull tiles.
+ */
+ spu.read_depth_stencil = (spu.depth_stencil_alpha.depth.enabled || spu.depth_stencil_alpha.stencil[0].enabled);
+
+ /* If we're forcing the fallback code to be used (for debug purposes),
+ * install that. Otherwise install the incoming SPU code.
+ */
+ if ((spu.init.debug_flags & CELL_DEBUG_FRAGMENT_OP_FALLBACK) != 0) {
+ static unsigned int warned = 0;
+ if (!warned) {
+ fprintf(stderr, "Cell Warning: using fallback per-fragment code\n");
+ warned = 1;
+ }
+ /* The following two lines aren't really necessary if you
+ * know the debug flags won't change during a run, and if you
+ * know that the function pointers are initialized correctly.
+ * We set them here to allow a person to change the debug
+ * flags during a run (from inside a debugger).
+ */
+ spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops;
+ spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops;
+ return;
+ }
+
+ /* Make sure the SPU code buffer is large enough to hold the incoming code.
+ * Note that we *don't* use align_malloc() and align_free(), because
+ * those utility functions are *not* available in SPU code.
+ * */
+ if (spu.fragment_ops_code_size < fops->total_code_size) {
+ if (spu.fragment_ops_code != NULL) {
+ free(spu.fragment_ops_code);
+ }
+ spu.fragment_ops_code_size = fops->total_code_size;
+ spu.fragment_ops_code = malloc(fops->total_code_size);
+ if (spu.fragment_ops_code == NULL) {
+ /* Whoops. */
+ fprintf(stderr, "CELL Warning: failed to allocate fragment ops code (%d bytes) - using fallback\n", fops->total_code_size);
+ spu.fragment_ops_code = NULL;
+ spu.fragment_ops_code_size = 0;
+ spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops;
+ spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops;
+ return;
+ }
+ }
+
+ /* Copy the SPU code from the command buffer to the spu buffer */
+ memcpy(spu.fragment_ops_code, fops->code, fops->total_code_size);
+
+ /* Set the pointers for the front-facing and back-facing fragments
+ * to the specified offsets within the code. Note that if the
+ * front-facing and back-facing code are the same, they'll have
+ * the same offset.
+ */
+ spu.fragment_ops[CELL_FACING_FRONT] = (spu_fragment_ops_func) &spu.fragment_ops_code[fops->front_code_index];
+ spu.fragment_ops[CELL_FACING_BACK] = (spu_fragment_ops_func) &spu.fragment_ops_code[fops->back_code_index];
+}
+
+static void
+cmd_state_fragment_program(const struct cell_command_fragment_program *fp)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_PROGRAM\n");
+ /* 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 uint
+cmd_state_fs_constants(const qword *buffer, uint pos)
+{
+ const uint num_const = spu_extract((vector unsigned int)buffer[pos+1], 0);
+ const float *constants = (const float *) &buffer[pos+2];
+ uint i;
+
+ D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FS_CONSTANTS (%u)\n", num_const);
+
+ /* Expand each float to float[4] for SOA execution */
+ for (i = 0; i < num_const; i++) {
+ D_PRINTF(CELL_DEBUG_CMD, " const[%u] = %f\n", i, constants[i]);
+ spu.constants[i] = spu_splats(constants[i]);
+ }
+
+ /* return new buffer pos (in 16-byte words) */
+ return pos + 2 + (ROUNDUP16(num_const * sizeof(float)) / 16);
+}
+
+
+static void
+cmd_state_framebuffer(const struct cell_command_framebuffer *cmd)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n",
+ cmd->width,
+ cmd->height,
+ cmd->color_start,
+ cmd->color_format,
+ cmd->depth_format);
+
+ ASSERT_ALIGN16(cmd->color_start);
+ ASSERT_ALIGN16(cmd->depth_start);
+
+ spu.fb.color_start = cmd->color_start;
+ spu.fb.depth_start = cmd->depth_start;
+ spu.fb.color_format = cmd->color_format;
+ spu.fb.depth_format = cmd->depth_format;
+ spu.fb.width = cmd->width;
+ spu.fb.height = cmd->height;
+ spu.fb.width_tiles = (spu.fb.width + TILE_SIZE - 1) / TILE_SIZE;
+ spu.fb.height_tiles = (spu.fb.height + TILE_SIZE - 1) / TILE_SIZE;
+
+ switch (spu.fb.depth_format) {
+ case PIPE_FORMAT_Z32_UNORM:
+ spu.fb.zsize = 4;
+ spu.fb.zscale = (float) 0xffffffffu;
+ 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;
+ case PIPE_FORMAT_Z16_UNORM:
+ spu.fb.zsize = 2;
+ spu.fb.zscale = (float) 0xffffu;
+ break;
+ default:
+ spu.fb.zsize = 0;
+ break;
+ }
+}
+
+
+/**
+ * Tex texture mask_s/t and scale_s/t fields depend on the texture size and
+ * sampler wrap modes.
+ */
+static void
+update_tex_masks(struct spu_texture *texture,
+ const struct pipe_sampler_state *sampler)
+{
+ uint i;
+
+ for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
+ int width = texture->level[i].width;
+ int height = texture->level[i].height;
+
+ if (sampler->wrap_s == PIPE_TEX_WRAP_REPEAT)
+ texture->level[i].mask_s = spu_splats(width - 1);
+ else
+ texture->level[i].mask_s = spu_splats(~0);
+
+ if (sampler->wrap_t == PIPE_TEX_WRAP_REPEAT)
+ texture->level[i].mask_t = spu_splats(height - 1);
+ else
+ texture->level[i].mask_t = spu_splats(~0);
+
+ if (sampler->normalized_coords) {
+ texture->level[i].scale_s = spu_splats((float) width);
+ texture->level[i].scale_t = spu_splats((float) height);
+ }
+ else {
+ texture->level[i].scale_s = spu_splats(1.0f);
+ texture->level[i].scale_t = spu_splats(1.0f);
+ }
+ }
+}
+
+
+static void
+cmd_state_sampler(const struct cell_command_sampler *sampler)
+{
+ uint unit = sampler->unit;
+
+ D_PRINTF(CELL_DEBUG_CMD, "SAMPLER [%u]\n", unit);
+
+ spu.sampler[unit] = sampler->state;
+
+ switch (spu.sampler[unit].min_img_filter) {
+ case PIPE_TEX_FILTER_LINEAR:
+ spu.min_sample_texture_2d[unit] = sample_texture_2d_bilinear;
+ break;
+ case PIPE_TEX_FILTER_ANISO:
+ /* fall-through, for now */
+ case PIPE_TEX_FILTER_NEAREST:
+ spu.min_sample_texture_2d[unit] = sample_texture_2d_nearest;
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ switch (spu.sampler[sampler->unit].mag_img_filter) {
+ case PIPE_TEX_FILTER_LINEAR:
+ spu.mag_sample_texture_2d[unit] = sample_texture_2d_bilinear;
+ break;
+ case PIPE_TEX_FILTER_ANISO:
+ /* fall-through, for now */
+ case PIPE_TEX_FILTER_NEAREST:
+ spu.mag_sample_texture_2d[unit] = sample_texture_2d_nearest;
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ switch (spu.sampler[sampler->unit].min_mip_filter) {
+ case PIPE_TEX_MIPFILTER_NEAREST:
+ case PIPE_TEX_MIPFILTER_LINEAR:
+ spu.sample_texture_2d[unit] = sample_texture_2d_lod;
+ break;
+ case PIPE_TEX_MIPFILTER_NONE:
+ spu.sample_texture_2d[unit] = spu.mag_sample_texture_2d[unit];
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ update_tex_masks(&spu.texture[unit], &spu.sampler[unit]);
+}
+
+
+static void
+cmd_state_texture(const struct cell_command_texture *texture)
+{
+ const uint unit = texture->unit;
+ uint i;
+
+ D_PRINTF(CELL_DEBUG_CMD, "TEXTURE [%u]\n", texture->unit);
+
+ spu.texture[unit].max_level = 0;
+ spu.texture[unit].target = texture->target;
+
+ for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
+ uint width = texture->width[i];
+ uint height = texture->height[i];
+ uint depth = texture->depth[i];
+
+ D_PRINTF(CELL_DEBUG_CMD, " LEVEL %u: at %p size[0] %u x %u\n", i,
+ texture->start[i], texture->width[i], texture->height[i]);
+
+ spu.texture[unit].level[i].start = texture->start[i];
+ spu.texture[unit].level[i].width = width;
+ spu.texture[unit].level[i].height = height;
+ spu.texture[unit].level[i].depth = depth;
+
+ spu.texture[unit].level[i].tiles_per_row =
+ (width + TILE_SIZE - 1) / TILE_SIZE;
+
+ spu.texture[unit].level[i].bytes_per_image =
+ 4 * align(width, TILE_SIZE) * align(height, TILE_SIZE) * depth;
+
+ spu.texture[unit].level[i].max_s = spu_splats((int) width - 1);
+ spu.texture[unit].level[i].max_t = spu_splats((int) height - 1);
+
+ if (texture->start[i])
+ spu.texture[unit].max_level = i;
+ }
+
+ update_tex_masks(&spu.texture[unit], &spu.sampler[unit]);
+}
+
+
+static void
+cmd_state_vertex_info(const struct vertex_info *vinfo)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs);
+ ASSERT(vinfo->num_attribs >= 1);
+ ASSERT(vinfo->num_attribs <= 8);
+ memcpy(&spu.vertex_info, vinfo, sizeof(*vinfo));
+}
+
+
+static void
+cmd_state_vs_array_info(const struct cell_array_info *vs_info)
+{
+ const unsigned attr = vs_info->attr;
+
+ ASSERT(attr < PIPE_MAX_ATTRIBS);
+ draw.vertex_fetch.src_ptr[attr] = vs_info->base;
+ draw.vertex_fetch.pitch[attr] = vs_info->pitch;
+ draw.vertex_fetch.size[attr] = vs_info->size;
+ draw.vertex_fetch.code_offset[attr] = vs_info->function_offset;
+ draw.vertex_fetch.dirty = 1;
+}
+
+
+static void
+cmd_state_attrib_fetch(const struct cell_attribute_fetch_code *code)
+{
+ mfc_get(attribute_fetch_code_buffer,
+ (unsigned int) code->base, /* src */
+ code->size,
+ TAG_BATCH_BUFFER,
+ 0, /* tid */
+ 0 /* rid */);
+ wait_on_mask(1 << TAG_BATCH_BUFFER);
+
+ draw.vertex_fetch.code = attribute_fetch_code_buffer;
+}
+
+
+static void
+cmd_finish(void)
+{
+ D_PRINTF(CELL_DEBUG_CMD, "FINISH\n");
+ really_clear_tiles(0);
+ /* wait for all outstanding DMAs to finish */
+ mfc_write_tag_mask(~0);
+ mfc_read_tag_status_all();
+ /* send mbox message to PPU */
+ spu_write_out_mbox(CELL_CMD_FINISH);
+}
+
+
+/**
+ * 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
+cmd_batch(uint opcode)
+{
+ const uint buf = (opcode >> 8) & 0xff;
+ uint size = (opcode >> 16);
+ qword buffer[CELL_BUFFER_SIZE / 16] ALIGN16_ATTRIB;
+ const unsigned usize = ROUNDUP16(size) / sizeof(buffer[0]);
+ uint pos;
+
+ D_PRINTF(CELL_DEBUG_CMD, "BATCH buffer %u, len %u, from %p\n",
+ buf, size, spu.init.buffers[buf]);
+
+ ASSERT((opcode & CELL_CMD_OPCODE_MASK) == CELL_CMD_BATCH);
+
+ ASSERT_ALIGN16(spu.init.buffers[buf]);
+
+ size = ROUNDUP16(size);
+
+ ASSERT_ALIGN16(spu.init.buffers[buf]);
+
+ mfc_get(buffer, /* dest */
+ (unsigned int) spu.init.buffers[buf], /* src */
+ size,
+ TAG_BATCH_BUFFER,
+ 0, /* tid */
+ 0 /* rid */);
+ wait_on_mask(1 << TAG_BATCH_BUFFER);
+
+ /* Tell PPU we're done copying the buffer to local store */
+ D_PRINTF(CELL_DEBUG_CMD, "release batch buf %u\n", buf);
+ release_buffer(buf);
+
+ /*
+ * Loop over commands in the batch buffer
+ */
+ for (pos = 0; pos < usize; /* no incr */) {
+ switch (si_to_uint(buffer[pos])) {
+ /*
+ * rendering commands
+ */
+ case CELL_CMD_CLEAR_SURFACE:
+ {
+ struct cell_command_clear_surface *clr
+ = (struct cell_command_clear_surface *) &buffer[pos];
+ cmd_clear_surface(clr);
+ pos += sizeof(*clr) / 16;
+ }
+ break;
+ case CELL_CMD_RENDER:
+ {
+ struct cell_command_render *render
+ = (struct cell_command_render *) &buffer[pos];
+ uint pos_incr;
+ cmd_render(render, &pos_incr);
+ pos += ((pos_incr+1)&~1) / 2; // should 'fix' cmd_render return
+ }
+ break;
+ /*
+ * state-update commands
+ */
+ case CELL_CMD_STATE_FRAMEBUFFER:
+ {
+ struct cell_command_framebuffer *fb
+ = (struct cell_command_framebuffer *) &buffer[pos];
+ cmd_state_framebuffer(fb);
+ pos += sizeof(*fb) / 16;
+ }
+ 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);
+ /* This is a variant-sized command */
+ pos += ROUNDUP16(sizeof(*fops) + fops->total_code_size) / 16;
+ }
+ 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) / 16;
+ }
+ break;
+ case CELL_CMD_STATE_FS_CONSTANTS:
+ pos = cmd_state_fs_constants(buffer, pos);
+ break;
+ case CELL_CMD_STATE_RASTERIZER:
+ {
+ struct cell_command_rasterizer *rast =
+ (struct cell_command_rasterizer *) &buffer[pos];
+ spu.rasterizer = rast->rasterizer;
+ pos += sizeof(*rast) / 16;
+ }
+ break;
+ case CELL_CMD_STATE_SAMPLER:
+ {
+ struct cell_command_sampler *sampler
+ = (struct cell_command_sampler *) &buffer[pos];
+ cmd_state_sampler(sampler);
+ pos += sizeof(*sampler) / 16;
+ }
+ break;
+ case CELL_CMD_STATE_TEXTURE:
+ {
+ struct cell_command_texture *texture
+ = (struct cell_command_texture *) &buffer[pos];
+ cmd_state_texture(texture);
+ pos += sizeof(*texture) / 16;
+ }
+ break;
+ case CELL_CMD_STATE_VERTEX_INFO:
+ cmd_state_vertex_info((struct vertex_info *) &buffer[pos+1]);
+ pos += 1 + ROUNDUP16(sizeof(struct vertex_info)) / 16;
+ break;
+ case CELL_CMD_STATE_VIEWPORT:
+ (void) memcpy(& draw.viewport, &buffer[pos+1],
+ sizeof(struct pipe_viewport_state));
+ pos += 1 + ROUNDUP16(sizeof(struct pipe_viewport_state)) / 16;
+ break;
+ case CELL_CMD_STATE_UNIFORMS:
+ draw.constants = (const float (*)[4]) (uintptr_t)spu_extract((vector unsigned int)buffer[pos+1],0);
+ pos += 2;
+ break;
+ case CELL_CMD_STATE_VS_ARRAY_INFO:
+ cmd_state_vs_array_info((struct cell_array_info *) &buffer[pos+1]);
+ pos += 1 + ROUNDUP16(sizeof(struct cell_array_info)) / 16;
+ break;
+ case CELL_CMD_STATE_BIND_VS:
+#if 0
+ spu_bind_vertex_shader(&draw,
+ (struct cell_shader_info *) &buffer[pos+1]);
+#endif
+ pos += 1 + ROUNDUP16(sizeof(struct cell_shader_info)) / 16;
+ break;
+ case CELL_CMD_STATE_ATTRIB_FETCH:
+ cmd_state_attrib_fetch((struct cell_attribute_fetch_code *)
+ &buffer[pos+1]);
+ pos += 1 + ROUNDUP16(sizeof(struct cell_attribute_fetch_code)) / 16;
+ break;
+ /*
+ * misc commands
+ */
+ case CELL_CMD_FINISH:
+ cmd_finish();
+ pos += 1;
+ break;
+ case CELL_CMD_FENCE:
+ {
+ struct cell_command_fence *fence_cmd =
+ (struct cell_command_fence *) &buffer[pos];
+ cmd_fence(fence_cmd);
+ pos += sizeof(*fence_cmd) / 16;
+ }
+ break;
+ case CELL_CMD_RELEASE_VERTS:
+ {
+ struct cell_command_release_verts *release
+ = (struct cell_command_release_verts *) &buffer[pos];
+ cmd_release_verts(release);
+ pos += sizeof(*release) / 16;
+ }
+ break;
+ case CELL_CMD_FLUSH_BUFFER_RANGE: {
+ struct cell_buffer_range *br = (struct cell_buffer_range *)
+ &buffer[pos+1];
+
+ spu_dcache_mark_dirty((unsigned) br->base, br->size);
+ pos += 1 + ROUNDUP16(sizeof(struct cell_buffer_range)) / 16;
+ break;
+ }
+ default:
+ printf("SPU %u: bad opcode: 0x%x\n", spu.init.id, si_to_uint(buffer[pos]));
+ ASSERT(0);
+ break;
+ }
+ }
+
+ D_PRINTF(CELL_DEBUG_CMD, "BATCH complete\n");
+}
+
+
+#define PERF 0
+
+
+/**
+ * Main loop for SPEs: Get a command, execute it, repeat.
+ */
+void
+command_loop(void)
+{
+ int exitFlag = 0;
+ uint t0, t1;
+
+ D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n");
+
+ while (!exitFlag) {
+ unsigned opcode;
+
+ D_PRINTF(CELL_DEBUG_CMD, "Wait for cmd...\n");
+
+ if (PERF)
+ spu_write_decrementer(~0);
+
+ /* read/wait from mailbox */
+ opcode = (unsigned int) spu_read_in_mbox();
+ D_PRINTF(CELL_DEBUG_CMD, "got cmd 0x%x\n", opcode);
+
+ if (PERF)
+ t0 = spu_read_decrementer();
+
+ switch (opcode & CELL_CMD_OPCODE_MASK) {
+ case CELL_CMD_EXIT:
+ D_PRINTF(CELL_DEBUG_CMD, "EXIT\n");
+ exitFlag = 1;
+ break;
+ case CELL_CMD_VS_EXECUTE:
+#if 0
+ spu_execute_vertex_shader(&draw, &cmd.vs);
+#endif
+ break;
+ case CELL_CMD_BATCH:
+ cmd_batch(opcode);
+ break;
+ default:
+ printf("Bad opcode 0x%x!\n", opcode & CELL_CMD_OPCODE_MASK);
+ }
+
+ if (PERF) {
+ t1 = spu_read_decrementer();
+ printf("wait mbox time: %gms batch time: %gms\n",
+ (~0u - t0) * spu.init.inv_timebase,
+ (t0 - t1) * spu.init.inv_timebase);
+ }
+ }
+
+ D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n");
+
+ if (spu.init.debug_flags & CELL_DEBUG_CACHE)
+ spu_dcache_report();
+}
+
+/* Initialize this module; we manage the fragment ops buffer here. */
+void
+spu_command_init(void)
+{
+ /* Install default/fallback fragment processing function.
+ * This will normally be overriden by a code-gen'd function
+ * unless CELL_FORCE_FRAGMENT_OPS_FALLBACK is set.
+ */
+ spu.fragment_ops[CELL_FACING_FRONT] = spu_fallback_fragment_ops;
+ spu.fragment_ops[CELL_FACING_BACK] = spu_fallback_fragment_ops;
+
+ /* Set up the basic empty buffer for code-gen'ed fragment ops */
+ spu.fragment_ops_code = NULL;
+ spu.fragment_ops_code_size = 0;
+}
+
+void
+spu_command_close(void)
+{
+ /* Deallocate the code-gen buffer for fragment ops, and reset the
+ * fragment ops functions to their initial setting (just to leave
+ * things in a good state).
+ */
+ if (spu.fragment_ops_code != NULL) {
+ free(spu.fragment_ops_code);
+ }
+ spu_command_init();
+}
diff --git a/src/gallium/drivers/cell/spu/spu_command.h b/src/gallium/drivers/cell/spu/spu_command.h
new file mode 100644
index 00000000000..83dcdade288
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_command.h
@@ -0,0 +1,35 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+extern void
+command_loop(void);
+
+extern void
+spu_command_init(void);
+
+extern void
+spu_command_close(void);
diff --git a/src/gallium/drivers/cell/spu/spu_dcache.c b/src/gallium/drivers/cell/spu/spu_dcache.c
new file mode 100644
index 00000000000..a6d67634fd8
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_dcache.c
@@ -0,0 +1,145 @@
+/*
+ * (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.
+ */
+
+#include "cell/common.h"
+#include "spu_main.h"
+#include "spu_dcache.h"
+
+#define CACHELINE_LOG2SIZE 7
+#define LINE_SIZE (1U << 7)
+#define ALIGN_MASK (~(LINE_SIZE - 1))
+
+#define CACHE_NAME data
+#define CACHED_TYPE qword
+#define CACHE_TYPE CACHE_TYPE_RO
+#define CACHE_SET_TAGID(set) (((set) & 0x03) + TAG_DCACHE0)
+#define CACHE_LOG2NNWAY 2
+#define CACHE_LOG2NSETS 6
+#ifdef DEBUG
+#define CACHE_STATS 1
+#endif
+#include <cache-api.h>
+
+/* Yes folks, this is ugly.
+ */
+#undef CACHE_NWAY
+#undef CACHE_NSETS
+#define CACHE_NAME data
+#define CACHE_NWAY 4
+#define CACHE_NSETS (1U << 6)
+
+
+/**
+ * Fetch between arbitrary number of bytes from an unaligned address
+ *
+ * \param dst Destination data buffer
+ * \param ea Main memory effective address of source data
+ * \param size Number of bytes to read
+ *
+ * \warning
+ * As is hinted by the type of the \c dst pointer, this function writes
+ * multiples of 16-bytes.
+ */
+void
+spu_dcache_fetch_unaligned(qword *dst, unsigned ea, unsigned size)
+{
+ const int shift = ea & 0x0f;
+ const unsigned read_size = ROUNDUP16(size + shift);
+ const unsigned last_read = ROUNDUP16(ea + size);
+ const qword *const last_write = dst + (ROUNDUP16(size) / 16);
+ unsigned i;
+
+
+ if (shift == 0) {
+ /* Data is already aligned. Fetch directly into the destination buffer.
+ */
+ for (i = 0; i < size; i += 16) {
+ *(dst++) = cache_rd(data, ea + i);
+ }
+ } else {
+ qword hi;
+
+
+ /* Please exercise extreme caution when modifying this code. This code
+ * must not read past the end of the page containing the source data,
+ * and it must not write more than ((size + 15) / 16) qwords to the
+ * destination buffer.
+ */
+ ea &= ~0x0f;
+ hi = cache_rd(data, ea);
+ for (i = 16; i < read_size; i += 16) {
+ qword lo = cache_rd(data, ea + i);
+
+ *(dst++) = si_or((qword) spu_slqwbyte(hi, shift),
+ (qword) spu_rlmaskqwbyte(lo, shift - 16));
+ hi = lo;
+ }
+
+ if (dst != last_write) {
+ *(dst++) = si_or((qword) spu_slqwbyte(hi, shift), si_il(0));
+ }
+ }
+
+ ASSERT((ea + i) == last_read);
+ ASSERT(dst == last_write);
+}
+
+
+/**
+ * Notify the cache that a range of main memory may have been modified
+ */
+void
+spu_dcache_mark_dirty(unsigned ea, unsigned size)
+{
+ unsigned i;
+ const unsigned aligned_start = (ea & ALIGN_MASK);
+ const unsigned aligned_end = (ea + size + (LINE_SIZE - 1))
+ & ALIGN_MASK;
+
+
+ for (i = 0; i < (CACHE_NWAY * CACHE_NSETS); i++) {
+ const unsigned entry = __cache_dir[i];
+ const unsigned addr = entry & ~0x0f;
+
+ __cache_dir[i] = ((addr >= aligned_start) && (addr < aligned_end))
+ ? (entry & ~CACHELINE_VALID) : entry;
+ }
+}
+
+
+/**
+ * Print cache utilization report
+ */
+void
+spu_dcache_report(void)
+{
+#ifdef CACHE_STATS
+ if (spu.init.id == 0) {
+ printf("SPU 0: Texture cache report:\n");
+ cache_pr_stats(data);
+ }
+#endif
+}
+
+
diff --git a/src/gallium/drivers/cell/spu/spu_dcache.h b/src/gallium/drivers/cell/spu/spu_dcache.h
new file mode 100644
index 00000000000..39a19eb31b5
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_dcache.h
@@ -0,0 +1,37 @@
+/*
+ * (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_DCACHE_H
+#define SPU_DCACHE_H
+
+extern void
+spu_dcache_fetch_unaligned(qword *dst, unsigned ea, unsigned size);
+
+extern void
+spu_dcache_mark_dirty(unsigned ea, unsigned size);
+
+extern void
+spu_dcache_report(void);
+
+#endif /* SPU_DCACHE_H */
diff --git a/src/gallium/drivers/cell/spu/spu_exec.c b/src/gallium/drivers/cell/spu/spu_exec.c
new file mode 100644
index 00000000000..e27df2dfb38
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_exec.c
@@ -0,0 +1,1933 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+/**
+ * TGSI interpretor/executor.
+ *
+ * Flow control information:
+ *
+ * Since we operate on 'quads' (4 pixels or 4 vertices in parallel)
+ * flow control statements (IF/ELSE/ENDIF, LOOP/ENDLOOP) require special
+ * care since a condition may be true for some quad components but false
+ * for other components.
+ *
+ * We basically execute all statements (even if they're in the part of
+ * an IF/ELSE clause that's "not taken") and use a special mask to
+ * control writing to destination registers. This is the ExecMask.
+ * See store_dest().
+ *
+ * The ExecMask is computed from three other masks (CondMask, LoopMask and
+ * ContMask) which are controlled by the flow control instructions (namely:
+ * (IF/ELSE/ENDIF, LOOP/ENDLOOP and CONT).
+ *
+ *
+ * Authors:
+ * Michal Krol
+ * Brian Paul
+ */
+
+#include <transpose_matrix4x4.h>
+#include <simdmath/ceilf4.h>
+#include <simdmath/cosf4.h>
+#include <simdmath/divf4.h>
+#include <simdmath/floorf4.h>
+#include <simdmath/log2f4.h>
+#include <simdmath/powf4.h>
+#include <simdmath/sinf4.h>
+#include <simdmath/sqrtf4.h>
+#include <simdmath/truncf4.h>
+
+#include "pipe/p_compiler.h"
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "tgsi/tgsi_parse.h"
+#include "tgsi/tgsi_util.h"
+#include "spu_exec.h"
+#include "spu_main.h"
+#include "spu_vertex_shader.h"
+#include "spu_dcache.h"
+#include "cell/common.h"
+
+#define TILE_TOP_LEFT 0
+#define TILE_TOP_RIGHT 1
+#define TILE_BOTTOM_LEFT 2
+#define TILE_BOTTOM_RIGHT 3
+
+/*
+ * Shorthand locations of various utility registers (_I = Index, _C = Channel)
+ */
+#define TEMP_0_I TGSI_EXEC_TEMP_00000000_I
+#define TEMP_0_C TGSI_EXEC_TEMP_00000000_C
+#define TEMP_7F_I TGSI_EXEC_TEMP_7FFFFFFF_I
+#define TEMP_7F_C TGSI_EXEC_TEMP_7FFFFFFF_C
+#define TEMP_80_I TGSI_EXEC_TEMP_80000000_I
+#define TEMP_80_C TGSI_EXEC_TEMP_80000000_C
+#define TEMP_FF_I TGSI_EXEC_TEMP_FFFFFFFF_I
+#define TEMP_FF_C TGSI_EXEC_TEMP_FFFFFFFF_C
+#define TEMP_1_I TGSI_EXEC_TEMP_ONE_I
+#define TEMP_1_C TGSI_EXEC_TEMP_ONE_C
+#define TEMP_2_I TGSI_EXEC_TEMP_TWO_I
+#define TEMP_2_C TGSI_EXEC_TEMP_TWO_C
+#define TEMP_128_I TGSI_EXEC_TEMP_128_I
+#define TEMP_128_C TGSI_EXEC_TEMP_128_C
+#define TEMP_M128_I TGSI_EXEC_TEMP_MINUS_128_I
+#define TEMP_M128_C TGSI_EXEC_TEMP_MINUS_128_C
+#define TEMP_KILMASK_I TGSI_EXEC_TEMP_KILMASK_I
+#define TEMP_KILMASK_C TGSI_EXEC_TEMP_KILMASK_C
+#define TEMP_OUTPUT_I TGSI_EXEC_TEMP_OUTPUT_I
+#define TEMP_OUTPUT_C TGSI_EXEC_TEMP_OUTPUT_C
+#define TEMP_PRIMITIVE_I TGSI_EXEC_TEMP_PRIMITIVE_I
+#define TEMP_PRIMITIVE_C TGSI_EXEC_TEMP_PRIMITIVE_C
+#define TEMP_R0 TGSI_EXEC_TEMP_R0
+
+#define FOR_EACH_CHANNEL(CHAN)\
+ for (CHAN = 0; CHAN < 4; CHAN++)
+
+#define IS_CHANNEL_ENABLED(INST, CHAN)\
+ ((INST).FullDstRegisters[0].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define IS_CHANNEL_ENABLED2(INST, CHAN)\
+ ((INST).FullDstRegisters[1].DstRegister.WriteMask & (1 << (CHAN)))
+
+#define FOR_EACH_ENABLED_CHANNEL(INST, CHAN)\
+ FOR_EACH_CHANNEL( CHAN )\
+ if (IS_CHANNEL_ENABLED( INST, CHAN ))
+
+#define FOR_EACH_ENABLED_CHANNEL2(INST, CHAN)\
+ FOR_EACH_CHANNEL( CHAN )\
+ if (IS_CHANNEL_ENABLED2( INST, CHAN ))
+
+
+/** The execution mask depends on the conditional mask and the loop mask */
+#define UPDATE_EXEC_MASK(MACH) \
+ MACH->ExecMask = MACH->CondMask & MACH->LoopMask & MACH->ContMask & MACH->FuncMask
+
+
+#define CHAN_X 0
+#define CHAN_Y 1
+#define CHAN_Z 2
+#define CHAN_W 3
+
+
+
+/**
+ * Initialize machine state by expanding tokens to full instructions,
+ * allocating temporary storage, setting up constants, etc.
+ * After this, we can call spu_exec_machine_run() many times.
+ */
+void
+spu_exec_machine_init(struct spu_exec_machine *mach,
+ uint numSamplers,
+ struct spu_sampler *samplers,
+ unsigned processor)
+{
+ const qword zero = si_il(0);
+ const qword not_zero = si_il(~0);
+
+ (void) numSamplers;
+ mach->Samplers = samplers;
+ mach->Processor = processor;
+ mach->Addrs = &mach->Temps[TGSI_EXEC_NUM_TEMPS];
+
+ /* Setup constants. */
+ mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q = zero;
+ mach->Temps[TEMP_FF_I].xyzw[TEMP_FF_C].q = not_zero;
+ mach->Temps[TEMP_7F_I].xyzw[TEMP_7F_C].q = si_shli(not_zero, -1);
+ mach->Temps[TEMP_80_I].xyzw[TEMP_80_C].q = si_shli(not_zero, 31);
+
+ mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].q = (qword) spu_splats(1.0f);
+ mach->Temps[TEMP_2_I].xyzw[TEMP_2_C].q = (qword) spu_splats(2.0f);
+ mach->Temps[TEMP_128_I].xyzw[TEMP_128_C].q = (qword) spu_splats(128.0f);
+ mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C].q = (qword) spu_splats(-128.0f);
+}
+
+
+static INLINE qword
+micro_abs(qword src)
+{
+ return si_rotmi(si_shli(src, 1), -1);
+}
+
+static INLINE qword
+micro_ceil(qword src)
+{
+ return (qword) _ceilf4((vec_float4) src);
+}
+
+static INLINE qword
+micro_cos(qword src)
+{
+ return (qword) _cosf4((vec_float4) src);
+}
+
+static const qword br_shuf = {
+ TILE_BOTTOM_RIGHT + 0, TILE_BOTTOM_RIGHT + 1,
+ TILE_BOTTOM_RIGHT + 2, TILE_BOTTOM_RIGHT + 3,
+ TILE_BOTTOM_RIGHT + 0, TILE_BOTTOM_RIGHT + 1,
+ TILE_BOTTOM_RIGHT + 2, TILE_BOTTOM_RIGHT + 3,
+ TILE_BOTTOM_RIGHT + 0, TILE_BOTTOM_RIGHT + 1,
+ TILE_BOTTOM_RIGHT + 2, TILE_BOTTOM_RIGHT + 3,
+ TILE_BOTTOM_RIGHT + 0, TILE_BOTTOM_RIGHT + 1,
+ TILE_BOTTOM_RIGHT + 2, TILE_BOTTOM_RIGHT + 3,
+};
+
+static const qword bl_shuf = {
+ TILE_BOTTOM_LEFT + 0, TILE_BOTTOM_LEFT + 1,
+ TILE_BOTTOM_LEFT + 2, TILE_BOTTOM_LEFT + 3,
+ TILE_BOTTOM_LEFT + 0, TILE_BOTTOM_LEFT + 1,
+ TILE_BOTTOM_LEFT + 2, TILE_BOTTOM_LEFT + 3,
+ TILE_BOTTOM_LEFT + 0, TILE_BOTTOM_LEFT + 1,
+ TILE_BOTTOM_LEFT + 2, TILE_BOTTOM_LEFT + 3,
+ TILE_BOTTOM_LEFT + 0, TILE_BOTTOM_LEFT + 1,
+ TILE_BOTTOM_LEFT + 2, TILE_BOTTOM_LEFT + 3,
+};
+
+static const qword tl_shuf = {
+ TILE_TOP_LEFT + 0, TILE_TOP_LEFT + 1,
+ TILE_TOP_LEFT + 2, TILE_TOP_LEFT + 3,
+ TILE_TOP_LEFT + 0, TILE_TOP_LEFT + 1,
+ TILE_TOP_LEFT + 2, TILE_TOP_LEFT + 3,
+ TILE_TOP_LEFT + 0, TILE_TOP_LEFT + 1,
+ TILE_TOP_LEFT + 2, TILE_TOP_LEFT + 3,
+ TILE_TOP_LEFT + 0, TILE_TOP_LEFT + 1,
+ TILE_TOP_LEFT + 2, TILE_TOP_LEFT + 3,
+};
+
+static qword
+micro_ddx(qword src)
+{
+ qword bottom_right = si_shufb(src, src, br_shuf);
+ qword bottom_left = si_shufb(src, src, bl_shuf);
+
+ return si_fs(bottom_right, bottom_left);
+}
+
+static qword
+micro_ddy(qword src)
+{
+ qword top_left = si_shufb(src, src, tl_shuf);
+ qword bottom_left = si_shufb(src, src, bl_shuf);
+
+ return si_fs(top_left, bottom_left);
+}
+
+static INLINE qword
+micro_div(qword src0, qword src1)
+{
+ return (qword) _divf4((vec_float4) src0, (vec_float4) src1);
+}
+
+static qword
+micro_flr(qword src)
+{
+ return (qword) _floorf4((vec_float4) src);
+}
+
+static qword
+micro_frc(qword src)
+{
+ return si_fs(src, (qword) _floorf4((vec_float4) src));
+}
+
+static INLINE qword
+micro_ge(qword src0, qword src1)
+{
+ return si_or(si_fceq(src0, src1), si_fcgt(src0, src1));
+}
+
+static qword
+micro_lg2(qword src)
+{
+ return (qword) _log2f4((vec_float4) src);
+}
+
+static INLINE qword
+micro_lt(qword src0, qword src1)
+{
+ const qword tmp = si_or(si_fceq(src0, src1), si_fcgt(src0, src1));
+
+ return si_xori(tmp, 0xff);
+}
+
+static INLINE qword
+micro_max(qword src0, qword src1)
+{
+ return si_selb(src1, src0, si_fcgt(src0, src1));
+}
+
+static INLINE qword
+micro_min(qword src0, qword src1)
+{
+ return si_selb(src0, src1, si_fcgt(src0, src1));
+}
+
+static qword
+micro_neg(qword src)
+{
+ return si_xor(src, (qword) spu_splats(0x80000000));
+}
+
+static qword
+micro_set_sign(qword src)
+{
+ return si_or(src, (qword) spu_splats(0x80000000));
+}
+
+static qword
+micro_pow(qword src0, qword src1)
+{
+ return (qword) _powf4((vec_float4) src0, (vec_float4) src1);
+}
+
+static qword
+micro_rnd(qword src)
+{
+ const qword half = (qword) spu_splats(0.5f);
+
+ /* May be able to use _roundf4. There may be some difference, though.
+ */
+ return (qword) _floorf4((vec_float4) si_fa(src, half));
+}
+
+static INLINE qword
+micro_ishr(qword src0, qword src1)
+{
+ return si_rotma(src0, si_sfi(src1, 0));
+}
+
+static qword
+micro_trunc(qword src)
+{
+ return (qword) _truncf4((vec_float4) src);
+}
+
+static qword
+micro_sin(qword src)
+{
+ return (qword) _sinf4((vec_float4) src);
+}
+
+static INLINE qword
+micro_sqrt(qword src)
+{
+ return (qword) _sqrtf4((vec_float4) src);
+}
+
+static void
+fetch_src_file_channel(
+ const struct spu_exec_machine *mach,
+ const uint file,
+ const uint swizzle,
+ const union spu_exec_channel *index,
+ union spu_exec_channel *chan )
+{
+ switch( swizzle ) {
+ case TGSI_EXTSWIZZLE_X:
+ case TGSI_EXTSWIZZLE_Y:
+ case TGSI_EXTSWIZZLE_Z:
+ case TGSI_EXTSWIZZLE_W:
+ switch( file ) {
+ case TGSI_FILE_CONSTANT: {
+ unsigned i;
+
+ for (i = 0; i < 4; i++) {
+ const float *ptr = mach->Consts[index->i[i]];
+ float tmp[4];
+
+ spu_dcache_fetch_unaligned((qword *) tmp,
+ (uintptr_t)(ptr + swizzle),
+ sizeof(float));
+
+ chan->f[i] = tmp[0];
+ }
+ break;
+ }
+
+ case TGSI_FILE_INPUT:
+ chan->u[0] = mach->Inputs[index->i[0]].xyzw[swizzle].u[0];
+ chan->u[1] = mach->Inputs[index->i[1]].xyzw[swizzle].u[1];
+ chan->u[2] = mach->Inputs[index->i[2]].xyzw[swizzle].u[2];
+ chan->u[3] = mach->Inputs[index->i[3]].xyzw[swizzle].u[3];
+ break;
+
+ case TGSI_FILE_TEMPORARY:
+ chan->u[0] = mach->Temps[index->i[0]].xyzw[swizzle].u[0];
+ chan->u[1] = mach->Temps[index->i[1]].xyzw[swizzle].u[1];
+ chan->u[2] = mach->Temps[index->i[2]].xyzw[swizzle].u[2];
+ chan->u[3] = mach->Temps[index->i[3]].xyzw[swizzle].u[3];
+ break;
+
+ case TGSI_FILE_IMMEDIATE:
+ ASSERT( index->i[0] < (int) mach->ImmLimit );
+ ASSERT( index->i[1] < (int) mach->ImmLimit );
+ ASSERT( index->i[2] < (int) mach->ImmLimit );
+ ASSERT( index->i[3] < (int) mach->ImmLimit );
+
+ chan->f[0] = mach->Imms[index->i[0]][swizzle];
+ chan->f[1] = mach->Imms[index->i[1]][swizzle];
+ chan->f[2] = mach->Imms[index->i[2]][swizzle];
+ chan->f[3] = mach->Imms[index->i[3]][swizzle];
+ break;
+
+ case TGSI_FILE_ADDRESS:
+ chan->u[0] = mach->Addrs[index->i[0]].xyzw[swizzle].u[0];
+ chan->u[1] = mach->Addrs[index->i[1]].xyzw[swizzle].u[1];
+ chan->u[2] = mach->Addrs[index->i[2]].xyzw[swizzle].u[2];
+ chan->u[3] = mach->Addrs[index->i[3]].xyzw[swizzle].u[3];
+ break;
+
+ case TGSI_FILE_OUTPUT:
+ /* vertex/fragment output vars can be read too */
+ chan->u[0] = mach->Outputs[index->i[0]].xyzw[swizzle].u[0];
+ chan->u[1] = mach->Outputs[index->i[1]].xyzw[swizzle].u[1];
+ chan->u[2] = mach->Outputs[index->i[2]].xyzw[swizzle].u[2];
+ chan->u[3] = mach->Outputs[index->i[3]].xyzw[swizzle].u[3];
+ break;
+
+ default:
+ ASSERT( 0 );
+ }
+ break;
+
+ case TGSI_EXTSWIZZLE_ZERO:
+ *chan = mach->Temps[TEMP_0_I].xyzw[TEMP_0_C];
+ break;
+
+ case TGSI_EXTSWIZZLE_ONE:
+ *chan = mach->Temps[TEMP_1_I].xyzw[TEMP_1_C];
+ break;
+
+ default:
+ ASSERT( 0 );
+ }
+}
+
+static void
+fetch_source(
+ const struct spu_exec_machine *mach,
+ union spu_exec_channel *chan,
+ const struct tgsi_full_src_register *reg,
+ const uint chan_index )
+{
+ union spu_exec_channel index;
+ uint swizzle;
+
+ index.i[0] =
+ index.i[1] =
+ index.i[2] =
+ index.i[3] = reg->SrcRegister.Index;
+
+ if (reg->SrcRegister.Indirect) {
+ union spu_exec_channel index2;
+ union spu_exec_channel indir_index;
+
+ index2.i[0] =
+ index2.i[1] =
+ index2.i[2] =
+ index2.i[3] = reg->SrcRegisterInd.Index;
+
+ swizzle = tgsi_util_get_src_register_swizzle(&reg->SrcRegisterInd,
+ CHAN_X);
+ fetch_src_file_channel(
+ mach,
+ reg->SrcRegisterInd.File,
+ swizzle,
+ &index2,
+ &indir_index );
+
+ index.q = si_a(index.q, indir_index.q);
+ }
+
+ if( reg->SrcRegister.Dimension ) {
+ switch( reg->SrcRegister.File ) {
+ case TGSI_FILE_INPUT:
+ index.q = si_mpyi(index.q, 17);
+ break;
+ case TGSI_FILE_CONSTANT:
+ index.q = si_shli(index.q, 12);
+ break;
+ default:
+ ASSERT( 0 );
+ }
+
+ index.i[0] += reg->SrcRegisterDim.Index;
+ index.i[1] += reg->SrcRegisterDim.Index;
+ index.i[2] += reg->SrcRegisterDim.Index;
+ index.i[3] += reg->SrcRegisterDim.Index;
+
+ if (reg->SrcRegisterDim.Indirect) {
+ union spu_exec_channel index2;
+ union spu_exec_channel indir_index;
+
+ index2.i[0] =
+ index2.i[1] =
+ index2.i[2] =
+ index2.i[3] = reg->SrcRegisterDimInd.Index;
+
+ swizzle = tgsi_util_get_src_register_swizzle( &reg->SrcRegisterDimInd, CHAN_X );
+ fetch_src_file_channel(
+ mach,
+ reg->SrcRegisterDimInd.File,
+ swizzle,
+ &index2,
+ &indir_index );
+
+ index.q = si_a(index.q, indir_index.q);
+ }
+ }
+
+ swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+ fetch_src_file_channel(
+ mach,
+ reg->SrcRegister.File,
+ swizzle,
+ &index,
+ chan );
+
+ switch (tgsi_util_get_full_src_register_sign_mode( reg, chan_index )) {
+ case TGSI_UTIL_SIGN_CLEAR:
+ chan->q = micro_abs(chan->q);
+ break;
+
+ case TGSI_UTIL_SIGN_SET:
+ chan->q = micro_set_sign(chan->q);
+ break;
+
+ case TGSI_UTIL_SIGN_TOGGLE:
+ chan->q = micro_neg(chan->q);
+ break;
+
+ case TGSI_UTIL_SIGN_KEEP:
+ break;
+ }
+
+ if (reg->SrcRegisterExtMod.Complement) {
+ chan->q = si_fs(mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].q, chan->q);
+ }
+}
+
+static void
+store_dest(
+ struct spu_exec_machine *mach,
+ const union spu_exec_channel *chan,
+ const struct tgsi_full_dst_register *reg,
+ const struct tgsi_full_instruction *inst,
+ uint chan_index )
+{
+ union spu_exec_channel *dst;
+
+ switch( reg->DstRegister.File ) {
+ case TGSI_FILE_NULL:
+ return;
+
+ case TGSI_FILE_OUTPUT:
+ dst = &mach->Outputs[mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0]
+ + reg->DstRegister.Index].xyzw[chan_index];
+ break;
+
+ case TGSI_FILE_TEMPORARY:
+ dst = &mach->Temps[reg->DstRegister.Index].xyzw[chan_index];
+ break;
+
+ case TGSI_FILE_ADDRESS:
+ dst = &mach->Addrs[reg->DstRegister.Index].xyzw[chan_index];
+ break;
+
+ default:
+ ASSERT( 0 );
+ return;
+ }
+
+ switch (inst->Instruction.Saturate)
+ {
+ case TGSI_SAT_NONE:
+ if (mach->ExecMask & 0x1)
+ dst->i[0] = chan->i[0];
+ if (mach->ExecMask & 0x2)
+ dst->i[1] = chan->i[1];
+ if (mach->ExecMask & 0x4)
+ dst->i[2] = chan->i[2];
+ if (mach->ExecMask & 0x8)
+ dst->i[3] = chan->i[3];
+ break;
+
+ case TGSI_SAT_ZERO_ONE:
+ /* XXX need to obey ExecMask here */
+ dst->q = micro_max(chan->q, mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q);
+ dst->q = micro_min(dst->q, mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].q);
+ break;
+
+ case TGSI_SAT_MINUS_PLUS_ONE:
+ ASSERT( 0 );
+ break;
+
+ default:
+ ASSERT( 0 );
+ }
+}
+
+#define FETCH(VAL,INDEX,CHAN)\
+ fetch_source (mach, VAL, &inst->FullSrcRegisters[INDEX], CHAN)
+
+#define STORE(VAL,INDEX,CHAN)\
+ store_dest (mach, VAL, &inst->FullDstRegisters[INDEX], inst, CHAN )
+
+
+/**
+ * Execute ARB-style KIL which is predicated by a src register.
+ * Kill fragment if any of the four values is less than zero.
+ */
+static void
+exec_kil(struct spu_exec_machine *mach,
+ const struct tgsi_full_instruction *inst)
+{
+ uint uniquemask;
+ uint chan_index;
+ uint kilmask = 0; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
+ union spu_exec_channel r[1];
+
+ /* This mask stores component bits that were already tested. Note that
+ * we test if the value is less than zero, so 1.0 and 0.0 need not to be
+ * tested. */
+ uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+
+ for (chan_index = 0; chan_index < 4; chan_index++)
+ {
+ uint swizzle;
+ uint i;
+
+ /* unswizzle channel */
+ swizzle = tgsi_util_get_full_src_register_extswizzle (
+ &inst->FullSrcRegisters[0],
+ chan_index);
+
+ /* check if the component has not been already tested */
+ if (uniquemask & (1 << swizzle))
+ continue;
+ uniquemask |= 1 << swizzle;
+
+ FETCH(&r[0], 0, chan_index);
+ for (i = 0; i < 4; i++)
+ if (r[0].f[i] < 0.0f)
+ kilmask |= 1 << i;
+ }
+
+ mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
+}
+
+/**
+ * Execute NVIDIA-style KIL which is predicated by a condition code.
+ * Kill fragment if the condition code is TRUE.
+ */
+static void
+exec_kilp(struct tgsi_exec_machine *mach,
+ const struct tgsi_full_instruction *inst)
+{
+ uint kilmask = 0; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
+
+ /* TODO: build kilmask from CC mask */
+
+ mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
+}
+
+/*
+ * Fetch a texel using STR texture coordinates.
+ */
+static void
+fetch_texel( struct spu_sampler *sampler,
+ const union spu_exec_channel *s,
+ const union spu_exec_channel *t,
+ const union spu_exec_channel *p,
+ float lodbias, /* XXX should be float[4] */
+ union spu_exec_channel *r,
+ union spu_exec_channel *g,
+ union spu_exec_channel *b,
+ union spu_exec_channel *a )
+{
+ qword rgba[4];
+ qword out[4];
+
+ sampler->get_samples(sampler, s->f, t->f, p->f, lodbias,
+ (float (*)[4]) rgba);
+
+ _transpose_matrix4x4((vec_float4 *) out, (vec_float4 *) rgba);
+ r->q = out[0];
+ g->q = out[1];
+ b->q = out[2];
+ a->q = out[3];
+}
+
+
+static void
+exec_tex(struct spu_exec_machine *mach,
+ const struct tgsi_full_instruction *inst,
+ boolean biasLod, boolean projected)
+{
+ const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index;
+ union spu_exec_channel r[8];
+ uint chan_index;
+ float lodBias;
+
+ /* printf("Sampler %u unit %u\n", sampler, unit); */
+
+ switch (inst->InstructionExtTexture.Texture) {
+ case TGSI_TEXTURE_1D:
+
+ FETCH(&r[0], 0, CHAN_X);
+
+ if (projected) {
+ FETCH(&r[1], 0, CHAN_W);
+ r[0].q = micro_div(r[0].q, r[1].q);
+ }
+
+ if (biasLod) {
+ FETCH(&r[1], 0, CHAN_W);
+ lodBias = r[2].f[0];
+ }
+ else
+ lodBias = 0.0;
+
+ fetch_texel(&mach->Samplers[unit],
+ &r[0], NULL, NULL, lodBias, /* S, T, P, BIAS */
+ &r[0], &r[1], &r[2], &r[3]); /* R, G, B, A */
+ break;
+
+ case TGSI_TEXTURE_2D:
+ case TGSI_TEXTURE_RECT:
+
+ FETCH(&r[0], 0, CHAN_X);
+ FETCH(&r[1], 0, CHAN_Y);
+ FETCH(&r[2], 0, CHAN_Z);
+
+ if (projected) {
+ FETCH(&r[3], 0, CHAN_W);
+ r[0].q = micro_div(r[0].q, r[3].q);
+ r[1].q = micro_div(r[1].q, r[3].q);
+ r[2].q = micro_div(r[2].q, r[3].q);
+ }
+
+ if (biasLod) {
+ FETCH(&r[3], 0, CHAN_W);
+ lodBias = r[3].f[0];
+ }
+ else
+ lodBias = 0.0;
+
+ fetch_texel(&mach->Samplers[unit],
+ &r[0], &r[1], &r[2], lodBias, /* inputs */
+ &r[0], &r[1], &r[2], &r[3]); /* outputs */
+ break;
+
+ case TGSI_TEXTURE_3D:
+ case TGSI_TEXTURE_CUBE:
+
+ FETCH(&r[0], 0, CHAN_X);
+ FETCH(&r[1], 0, CHAN_Y);
+ FETCH(&r[2], 0, CHAN_Z);
+
+ if (projected) {
+ FETCH(&r[3], 0, CHAN_W);
+ r[0].q = micro_div(r[0].q, r[3].q);
+ r[1].q = micro_div(r[1].q, r[3].q);
+ r[2].q = micro_div(r[2].q, r[3].q);
+ }
+
+ if (biasLod) {
+ FETCH(&r[3], 0, CHAN_W);
+ lodBias = r[3].f[0];
+ }
+ else
+ lodBias = 0.0;
+
+ fetch_texel(&mach->Samplers[unit],
+ &r[0], &r[1], &r[2], lodBias,
+ &r[0], &r[1], &r[2], &r[3]);
+ break;
+
+ default:
+ ASSERT (0);
+ }
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[chan_index], 0, chan_index );
+ }
+}
+
+
+
+static void
+constant_interpolation(
+ struct spu_exec_machine *mach,
+ unsigned attrib,
+ unsigned chan )
+{
+ unsigned i;
+
+ for( i = 0; i < QUAD_SIZE; i++ ) {
+ mach->Inputs[attrib].xyzw[chan].f[i] = mach->InterpCoefs[attrib].a0[chan];
+ }
+}
+
+static void
+linear_interpolation(
+ struct spu_exec_machine *mach,
+ unsigned attrib,
+ unsigned chan )
+{
+ const float x = mach->QuadPos.xyzw[0].f[0];
+ const float y = mach->QuadPos.xyzw[1].f[0];
+ const float dadx = mach->InterpCoefs[attrib].dadx[chan];
+ const float dady = mach->InterpCoefs[attrib].dady[chan];
+ const float a0 = mach->InterpCoefs[attrib].a0[chan] + dadx * x + dady * y;
+ mach->Inputs[attrib].xyzw[chan].f[0] = a0;
+ mach->Inputs[attrib].xyzw[chan].f[1] = a0 + dadx;
+ mach->Inputs[attrib].xyzw[chan].f[2] = a0 + dady;
+ mach->Inputs[attrib].xyzw[chan].f[3] = a0 + dadx + dady;
+}
+
+static void
+perspective_interpolation(
+ struct spu_exec_machine *mach,
+ unsigned attrib,
+ unsigned chan )
+{
+ const float x = mach->QuadPos.xyzw[0].f[0];
+ const float y = mach->QuadPos.xyzw[1].f[0];
+ const float dadx = mach->InterpCoefs[attrib].dadx[chan];
+ const float dady = mach->InterpCoefs[attrib].dady[chan];
+ const float a0 = mach->InterpCoefs[attrib].a0[chan] + dadx * x + dady * y;
+ const float *w = mach->QuadPos.xyzw[3].f;
+ /* divide by W here */
+ mach->Inputs[attrib].xyzw[chan].f[0] = a0 / w[0];
+ mach->Inputs[attrib].xyzw[chan].f[1] = (a0 + dadx) / w[1];
+ mach->Inputs[attrib].xyzw[chan].f[2] = (a0 + dady) / w[2];
+ mach->Inputs[attrib].xyzw[chan].f[3] = (a0 + dadx + dady) / w[3];
+}
+
+
+typedef void (* interpolation_func)(
+ struct spu_exec_machine *mach,
+ unsigned attrib,
+ unsigned chan );
+
+static void
+exec_declaration(struct spu_exec_machine *mach,
+ const struct tgsi_full_declaration *decl)
+{
+ if( mach->Processor == TGSI_PROCESSOR_FRAGMENT ) {
+ if( decl->Declaration.File == TGSI_FILE_INPUT ) {
+ unsigned first, last, mask;
+ interpolation_func interp;
+
+ first = decl->DeclarationRange.First;
+ last = decl->DeclarationRange.Last;
+ mask = decl->Declaration.UsageMask;
+
+ switch( decl->Declaration.Interpolate ) {
+ case TGSI_INTERPOLATE_CONSTANT:
+ interp = constant_interpolation;
+ break;
+
+ case TGSI_INTERPOLATE_LINEAR:
+ interp = linear_interpolation;
+ break;
+
+ case TGSI_INTERPOLATE_PERSPECTIVE:
+ interp = perspective_interpolation;
+ break;
+
+ default:
+ ASSERT( 0 );
+ }
+
+ if( mask == TGSI_WRITEMASK_XYZW ) {
+ unsigned i, j;
+
+ for( i = first; i <= last; i++ ) {
+ for( j = 0; j < NUM_CHANNELS; j++ ) {
+ interp( mach, i, j );
+ }
+ }
+ }
+ else {
+ unsigned i, j;
+
+ for( j = 0; j < NUM_CHANNELS; j++ ) {
+ if( mask & (1 << j) ) {
+ for( i = first; i <= last; i++ ) {
+ interp( mach, i, j );
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+static void
+exec_instruction(
+ struct spu_exec_machine *mach,
+ const struct tgsi_full_instruction *inst,
+ int *pc )
+{
+ uint chan_index;
+ union spu_exec_channel r[8];
+
+ (*pc)++;
+
+ switch (inst->Instruction.Opcode) {
+ case TGSI_OPCODE_ARL:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = si_cflts(r[0].q, 0);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_MOV:
+ case TGSI_OPCODE_SWZ:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_LIT:
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Y ) || IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+ FETCH( &r[0], 0, CHAN_X );
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+ r[0].q = micro_max(r[0].q, mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q);
+ STORE( &r[0], 0, CHAN_Y );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+ FETCH( &r[1], 0, CHAN_Y );
+ r[1].q = micro_max(r[1].q, mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q);
+
+ FETCH( &r[2], 0, CHAN_W );
+ r[2].q = micro_min(r[2].q, mach->Temps[TEMP_128_I].xyzw[TEMP_128_C].q);
+ r[2].q = micro_max(r[2].q, mach->Temps[TEMP_M128_I].xyzw[TEMP_M128_C].q);
+ r[1].q = micro_pow(r[1].q, r[2].q);
+
+ /* r0 = (r0 > 0.0) ? r1 : 0.0
+ */
+ r[0].q = si_fcgt(r[0].q, mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q);
+ r[0].q = si_selb(mach->Temps[TEMP_0_I].xyzw[TEMP_0_C].q, r[1].q,
+ r[0].q);
+ STORE( &r[0], 0, CHAN_Z );
+ }
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+ }
+ break;
+
+ case TGSI_OPCODE_RCP:
+ /* TGSI_OPCODE_RECIP */
+ FETCH( &r[0], 0, CHAN_X );
+ r[0].q = micro_div(mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].q, r[0].q);
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_RSQ:
+ /* TGSI_OPCODE_RECIPSQRT */
+ FETCH( &r[0], 0, CHAN_X );
+ r[0].q = micro_sqrt(r[0].q);
+ r[0].q = micro_div(mach->Temps[TEMP_1_I].xyzw[TEMP_1_C].q, r[0].q);
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_EXP:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_LOG:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_MUL:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index )
+ {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+
+ r[0].q = si_fm(r[0].q, r[1].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_ADD:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = si_fa(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_DP3:
+ /* TGSI_OPCODE_DOT3 */
+ FETCH( &r[0], 0, CHAN_X );
+ FETCH( &r[1], 1, CHAN_X );
+ r[0].q = si_fm(r[0].q, r[1].q);
+
+ FETCH( &r[1], 0, CHAN_Y );
+ FETCH( &r[2], 1, CHAN_Y );
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+
+ FETCH( &r[1], 0, CHAN_Z );
+ FETCH( &r[2], 1, CHAN_Z );
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_DP4:
+ /* TGSI_OPCODE_DOT4 */
+ FETCH(&r[0], 0, CHAN_X);
+ FETCH(&r[1], 1, CHAN_X);
+
+ r[0].q = si_fm(r[0].q, r[1].q);
+
+ FETCH(&r[1], 0, CHAN_Y);
+ FETCH(&r[2], 1, CHAN_Y);
+
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FETCH(&r[1], 0, CHAN_Z);
+ FETCH(&r[2], 1, CHAN_Z);
+
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FETCH(&r[1], 0, CHAN_W);
+ FETCH(&r[2], 1, CHAN_W);
+
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_DST:
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_X );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+ FETCH( &r[0], 0, CHAN_Y );
+ FETCH( &r[1], 1, CHAN_Y);
+ r[0].q = si_fm(r[0].q, r[1].q);
+ STORE( &r[0], 0, CHAN_Y );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+ FETCH( &r[0], 0, CHAN_Z );
+ STORE( &r[0], 0, CHAN_Z );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+ FETCH( &r[0], 1, CHAN_W );
+ STORE( &r[0], 0, CHAN_W );
+ }
+ break;
+
+ case TGSI_OPCODE_MIN:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+
+ r[0].q = micro_min(r[0].q, r[1].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_MAX:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+
+ r[0].q = micro_max(r[0].q, r[1].q);
+
+ STORE(&r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SLT:
+ /* TGSI_OPCODE_SETLT */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+
+ r[0].q = micro_ge(r[0].q, r[1].q);
+ r[0].q = si_xori(r[0].q, 0xff);
+
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SGE:
+ /* TGSI_OPCODE_SETGE */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = micro_ge(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_MAD:
+ /* TGSI_OPCODE_MADD */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ FETCH( &r[2], 2, chan_index );
+ r[0].q = si_fma(r[0].q, r[1].q, r[2].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SUB:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+
+ r[0].q = si_fs(r[0].q, r[1].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_LERP:
+ /* TGSI_OPCODE_LRP */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+ FETCH(&r[2], 2, chan_index);
+
+ r[1].q = si_fs(r[1].q, r[2].q);
+ r[0].q = si_fma(r[0].q, r[1].q, r[2].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_CND:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_CND0:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_DOT2ADD:
+ /* TGSI_OPCODE_DP2A */
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_INDEX:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_NEGATE:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_FRAC:
+ /* TGSI_OPCODE_FRC */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_frc(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_CLAMP:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_FLOOR:
+ /* TGSI_OPCODE_FLR */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_flr(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_ROUND:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_rnd(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_EXPBASE2:
+ /* TGSI_OPCODE_EX2 */
+ FETCH(&r[0], 0, CHAN_X);
+
+ r[0].q = micro_pow(mach->Temps[TEMP_2_I].xyzw[TEMP_2_C].q, r[0].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_LOGBASE2:
+ /* TGSI_OPCODE_LG2 */
+ FETCH( &r[0], 0, CHAN_X );
+ r[0].q = micro_lg2(r[0].q);
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_POWER:
+ /* TGSI_OPCODE_POW */
+ FETCH(&r[0], 0, CHAN_X);
+ FETCH(&r[1], 1, CHAN_X);
+
+ r[0].q = micro_pow(r[0].q, r[1].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_CROSSPRODUCT:
+ /* TGSI_OPCODE_XPD */
+ FETCH(&r[0], 0, CHAN_Y);
+ FETCH(&r[1], 1, CHAN_Z);
+ FETCH(&r[3], 0, CHAN_Z);
+ FETCH(&r[4], 1, CHAN_Y);
+
+ /* r2 = (r0 * r1) - (r3 * r5)
+ */
+ r[2].q = si_fm(r[3].q, r[5].q);
+ r[2].q = si_fms(r[0].q, r[1].q, r[2].q);
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
+ STORE( &r[2], 0, CHAN_X );
+ }
+
+ FETCH(&r[2], 1, CHAN_X);
+ FETCH(&r[5], 0, CHAN_X);
+
+ /* r3 = (r3 * r2) - (r1 * r5)
+ */
+ r[1].q = si_fm(r[1].q, r[5].q);
+ r[3].q = si_fms(r[3].q, r[2].q, r[1].q);
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
+ STORE( &r[3], 0, CHAN_Y );
+ }
+
+ /* r5 = (r5 * r4) - (r0 * r2)
+ */
+ r[0].q = si_fm(r[0].q, r[2].q);
+ r[5].q = si_fms(r[5].q, r[4].q, r[0].q);
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
+ STORE( &r[5], 0, CHAN_Z );
+ }
+
+ if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+ }
+ break;
+
+ case TGSI_OPCODE_MULTIPLYMATRIX:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_ABS:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+
+ r[0].q = micro_abs(r[0].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_RCC:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_DPH:
+ FETCH(&r[0], 0, CHAN_X);
+ FETCH(&r[1], 1, CHAN_X);
+
+ r[0].q = si_fm(r[0].q, r[1].q);
+
+ FETCH(&r[1], 0, CHAN_Y);
+ FETCH(&r[2], 1, CHAN_Y);
+
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FETCH(&r[1], 0, CHAN_Z);
+ FETCH(&r[2], 1, CHAN_Z);
+
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FETCH(&r[1], 1, CHAN_W);
+
+ r[0].q = si_fa(r[0].q, r[1].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_COS:
+ FETCH(&r[0], 0, CHAN_X);
+
+ r[0].q = micro_cos(r[0].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_DDX:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_ddx(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_DDY:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_ddy(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_KILP:
+ exec_kilp (mach, inst);
+ break;
+
+ case TGSI_OPCODE_KIL:
+ exec_kil (mach, inst);
+ break;
+
+ case TGSI_OPCODE_PK2H:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_PK2US:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_PK4B:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_PK4UB:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_RFL:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_SEQ:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+
+ r[0].q = si_fceq(r[0].q, r[1].q);
+
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SFL:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_SGT:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = si_fcgt(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SIN:
+ FETCH( &r[0], 0, CHAN_X );
+ r[0].q = micro_sin(r[0].q);
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SLE:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+
+ r[0].q = si_fcgt(r[0].q, r[1].q);
+ r[0].q = si_xori(r[0].q, 0xff);
+
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SNE:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+
+ r[0].q = si_fceq(r[0].q, r[1].q);
+ r[0].q = si_xori(r[0].q, 0xff);
+
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_STR:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_TEX:
+ /* simple texture lookup */
+ /* src[0] = texcoord */
+ /* src[1] = sampler unit */
+ exec_tex(mach, inst, FALSE, FALSE);
+ break;
+
+ case TGSI_OPCODE_TXB:
+ /* Texture lookup with lod bias */
+ /* src[0] = texcoord (src[0].w = load bias) */
+ /* src[1] = sampler unit */
+ exec_tex(mach, inst, TRUE, FALSE);
+ break;
+
+ case TGSI_OPCODE_TXD:
+ /* Texture lookup with explict partial derivatives */
+ /* src[0] = texcoord */
+ /* src[1] = d[strq]/dx */
+ /* src[2] = d[strq]/dy */
+ /* src[3] = sampler unit */
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_TXL:
+ /* Texture lookup with explit LOD */
+ /* src[0] = texcoord (src[0].w = load bias) */
+ /* src[1] = sampler unit */
+ exec_tex(mach, inst, TRUE, FALSE);
+ break;
+
+ case TGSI_OPCODE_TXP:
+ /* Texture lookup with projection */
+ /* src[0] = texcoord (src[0].w = projection) */
+ /* src[1] = sampler unit */
+ exec_tex(mach, inst, TRUE, TRUE);
+ break;
+
+ case TGSI_OPCODE_UP2H:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_UP2US:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_UP4B:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_UP4UB:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_X2D:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_ARA:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_ARR:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_BRA:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_CAL:
+ /* skip the call if no execution channels are enabled */
+ if (mach->ExecMask) {
+ /* do the call */
+
+ /* push the Cond, Loop, Cont stacks */
+ ASSERT(mach->CondStackTop < TGSI_EXEC_MAX_COND_NESTING);
+ mach->CondStack[mach->CondStackTop++] = mach->CondMask;
+ ASSERT(mach->LoopStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+ mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
+ ASSERT(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+ mach->ContStack[mach->ContStackTop++] = mach->ContMask;
+
+ ASSERT(mach->FuncStackTop < TGSI_EXEC_MAX_CALL_NESTING);
+ mach->FuncStack[mach->FuncStackTop++] = mach->FuncMask;
+
+ /* note that PC was already incremented above */
+ mach->CallStack[mach->CallStackTop++] = *pc;
+ *pc = inst->InstructionExtLabel.Label;
+ }
+ break;
+
+ case TGSI_OPCODE_RET:
+ mach->FuncMask &= ~mach->ExecMask;
+ UPDATE_EXEC_MASK(mach);
+
+ if (mach->ExecMask == 0x0) {
+ /* really return now (otherwise, keep executing */
+
+ if (mach->CallStackTop == 0) {
+ /* returning from main() */
+ *pc = -1;
+ return;
+ }
+ *pc = mach->CallStack[--mach->CallStackTop];
+
+ /* pop the Cond, Loop, Cont stacks */
+ ASSERT(mach->CondStackTop > 0);
+ mach->CondMask = mach->CondStack[--mach->CondStackTop];
+ ASSERT(mach->LoopStackTop > 0);
+ mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
+ ASSERT(mach->ContStackTop > 0);
+ mach->ContMask = mach->ContStack[--mach->ContStackTop];
+ ASSERT(mach->FuncStackTop > 0);
+ mach->FuncMask = mach->FuncStack[--mach->FuncStackTop];
+
+ UPDATE_EXEC_MASK(mach);
+ }
+ break;
+
+ case TGSI_OPCODE_SSG:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_CMP:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH(&r[0], 0, chan_index);
+ FETCH(&r[1], 1, chan_index);
+ FETCH(&r[2], 2, chan_index);
+
+ /* r0 = (r0 < 0.0) ? r1 : r2
+ */
+ r[3].q = si_xor(r[3].q, r[3].q);
+ r[0].q = micro_lt(r[0].q, r[3].q);
+ r[0].q = si_selb(r[1].q, r[2].q, r[0].q);
+
+ STORE(&r[0], 0, chan_index);
+ }
+ break;
+
+ case TGSI_OPCODE_SCS:
+ if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) || IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+ FETCH( &r[0], 0, CHAN_X );
+ }
+ if( IS_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
+ r[1].q = micro_cos(r[0].q);
+ STORE( &r[1], 0, CHAN_X );
+ }
+ if( IS_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+ r[1].q = micro_sin(r[0].q);
+ STORE( &r[1], 0, CHAN_Y );
+ }
+ if( IS_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+ STORE( &mach->Temps[TEMP_0_I].xyzw[TEMP_0_C], 0, CHAN_Z );
+ }
+ if( IS_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+ }
+ break;
+
+ case TGSI_OPCODE_NRM:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_DIV:
+ ASSERT( 0 );
+ break;
+
+ case TGSI_OPCODE_DP2:
+ FETCH( &r[0], 0, CHAN_X );
+ FETCH( &r[1], 1, CHAN_X );
+ r[0].q = si_fm(r[0].q, r[1].q);
+
+ FETCH( &r[1], 0, CHAN_Y );
+ FETCH( &r[2], 1, CHAN_Y );
+ r[0].q = si_fma(r[1].q, r[2].q, r[0].q);
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_IF:
+ /* push CondMask */
+ ASSERT(mach->CondStackTop < TGSI_EXEC_MAX_COND_NESTING);
+ mach->CondStack[mach->CondStackTop++] = mach->CondMask;
+ FETCH( &r[0], 0, CHAN_X );
+ /* update CondMask */
+ if( ! r[0].u[0] ) {
+ mach->CondMask &= ~0x1;
+ }
+ if( ! r[0].u[1] ) {
+ mach->CondMask &= ~0x2;
+ }
+ if( ! r[0].u[2] ) {
+ mach->CondMask &= ~0x4;
+ }
+ if( ! r[0].u[3] ) {
+ mach->CondMask &= ~0x8;
+ }
+ UPDATE_EXEC_MASK(mach);
+ /* Todo: If CondMask==0, jump to ELSE */
+ break;
+
+ case TGSI_OPCODE_ELSE:
+ /* invert CondMask wrt previous mask */
+ {
+ uint prevMask;
+ ASSERT(mach->CondStackTop > 0);
+ prevMask = mach->CondStack[mach->CondStackTop - 1];
+ mach->CondMask = ~mach->CondMask & prevMask;
+ UPDATE_EXEC_MASK(mach);
+ /* Todo: If CondMask==0, jump to ENDIF */
+ }
+ break;
+
+ case TGSI_OPCODE_ENDIF:
+ /* pop CondMask */
+ ASSERT(mach->CondStackTop > 0);
+ mach->CondMask = mach->CondStack[--mach->CondStackTop];
+ UPDATE_EXEC_MASK(mach);
+ break;
+
+ case TGSI_OPCODE_END:
+ /* halt execution */
+ *pc = -1;
+ break;
+
+ case TGSI_OPCODE_REP:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_ENDREP:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_PUSHA:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_POPA:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_CEIL:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_ceil(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_I2F:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = si_csflt(r[0].q, 0);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_NOT:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = si_xorbi(r[0].q, 0xff);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_TRUNC:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ r[0].q = micro_trunc(r[0].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SHL:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+
+ r[0].q = si_shl(r[0].q, r[1].q);
+
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SHR:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = micro_ishr(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_AND:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = si_and(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_OR:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = si_or(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_MOD:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_XOR:
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ FETCH( &r[1], 1, chan_index );
+ r[0].q = si_xor(r[0].q, r[1].q);
+ STORE( &r[0], 0, chan_index );
+ }
+ break;
+
+ case TGSI_OPCODE_SAD:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_TXF:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_TXQ:
+ ASSERT (0);
+ break;
+
+ case TGSI_OPCODE_EMIT:
+ mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] += 16;
+ mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]]++;
+ break;
+
+ case TGSI_OPCODE_ENDPRIM:
+ mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]++;
+ mach->Primitives[mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0]] = 0;
+ break;
+
+ case TGSI_OPCODE_LOOP:
+ /* fall-through (for now) */
+ case TGSI_OPCODE_BGNLOOP2:
+ /* push LoopMask and ContMasks */
+ ASSERT(mach->LoopStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+ mach->LoopStack[mach->LoopStackTop++] = mach->LoopMask;
+ ASSERT(mach->ContStackTop < TGSI_EXEC_MAX_LOOP_NESTING);
+ mach->ContStack[mach->ContStackTop++] = mach->ContMask;
+ break;
+
+ case TGSI_OPCODE_ENDLOOP:
+ /* fall-through (for now at least) */
+ case TGSI_OPCODE_ENDLOOP2:
+ /* Restore ContMask, but don't pop */
+ ASSERT(mach->ContStackTop > 0);
+ mach->ContMask = mach->ContStack[mach->ContStackTop - 1];
+ if (mach->LoopMask) {
+ /* repeat loop: jump to instruction just past BGNLOOP */
+ *pc = inst->InstructionExtLabel.Label + 1;
+ }
+ else {
+ /* exit loop: pop LoopMask */
+ ASSERT(mach->LoopStackTop > 0);
+ mach->LoopMask = mach->LoopStack[--mach->LoopStackTop];
+ /* pop ContMask */
+ ASSERT(mach->ContStackTop > 0);
+ mach->ContMask = mach->ContStack[--mach->ContStackTop];
+ }
+ UPDATE_EXEC_MASK(mach);
+ break;
+
+ case TGSI_OPCODE_BRK:
+ /* turn off loop channels for each enabled exec channel */
+ mach->LoopMask &= ~mach->ExecMask;
+ /* Todo: if mach->LoopMask == 0, jump to end of loop */
+ UPDATE_EXEC_MASK(mach);
+ break;
+
+ case TGSI_OPCODE_CONT:
+ /* turn off cont channels for each enabled exec channel */
+ mach->ContMask &= ~mach->ExecMask;
+ /* Todo: if mach->LoopMask == 0, jump to end of loop */
+ UPDATE_EXEC_MASK(mach);
+ break;
+
+ case TGSI_OPCODE_BGNSUB:
+ /* no-op */
+ break;
+
+ case TGSI_OPCODE_ENDSUB:
+ /* no-op */
+ break;
+
+ case TGSI_OPCODE_NOISE1:
+ ASSERT( 0 );
+ break;
+
+ case TGSI_OPCODE_NOISE2:
+ ASSERT( 0 );
+ break;
+
+ case TGSI_OPCODE_NOISE3:
+ ASSERT( 0 );
+ break;
+
+ case TGSI_OPCODE_NOISE4:
+ ASSERT( 0 );
+ break;
+
+ case TGSI_OPCODE_NOP:
+ break;
+
+ default:
+ ASSERT( 0 );
+ }
+}
+
+
+/**
+ * Run TGSI interpreter.
+ * \return bitmask of "alive" quad components
+ */
+uint
+spu_exec_machine_run( struct spu_exec_machine *mach )
+{
+ uint i;
+ int pc = 0;
+
+ mach->CondMask = 0xf;
+ mach->LoopMask = 0xf;
+ mach->ContMask = 0xf;
+ mach->FuncMask = 0xf;
+ mach->ExecMask = 0xf;
+
+ mach->CondStackTop = 0; /* temporarily subvert this ASSERTion */
+ ASSERT(mach->CondStackTop == 0);
+ ASSERT(mach->LoopStackTop == 0);
+ ASSERT(mach->ContStackTop == 0);
+ ASSERT(mach->CallStackTop == 0);
+
+ mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] = 0;
+ mach->Temps[TEMP_OUTPUT_I].xyzw[TEMP_OUTPUT_C].u[0] = 0;
+
+ if( mach->Processor == TGSI_PROCESSOR_GEOMETRY ) {
+ mach->Temps[TEMP_PRIMITIVE_I].xyzw[TEMP_PRIMITIVE_C].u[0] = 0;
+ mach->Primitives[0] = 0;
+ }
+
+
+ /* execute declarations (interpolants) */
+ if( mach->Processor == TGSI_PROCESSOR_FRAGMENT ) {
+ for (i = 0; i < mach->NumDeclarations; i++) {
+ union {
+ struct tgsi_full_declaration decl;
+ qword buffer[ROUNDUP16(sizeof(struct tgsi_full_declaration)) / 16];
+ } d ALIGN16_ATTRIB;
+ unsigned ea = (unsigned) (mach->Declarations + pc);
+
+ spu_dcache_fetch_unaligned(d.buffer, ea, sizeof(d.decl));
+
+ exec_declaration( mach, &d.decl );
+ }
+ }
+
+ /* execute instructions, until pc is set to -1 */
+ while (pc != -1) {
+ union {
+ struct tgsi_full_instruction inst;
+ qword buffer[ROUNDUP16(sizeof(struct tgsi_full_instruction)) / 16];
+ } i ALIGN16_ATTRIB;
+ unsigned ea = (unsigned) (mach->Instructions + pc);
+
+ spu_dcache_fetch_unaligned(i.buffer, ea, sizeof(i.inst));
+ exec_instruction( mach, & i.inst, &pc );
+ }
+
+#if 0
+ /* we scale from floats in [0,1] to Zbuffer ints in sp_quad_depth_test.c */
+ if (mach->Processor == TGSI_PROCESSOR_FRAGMENT) {
+ /*
+ * Scale back depth component.
+ */
+ for (i = 0; i < 4; i++)
+ mach->Outputs[0].xyzw[2].f[i] *= ctx->DrawBuffer->_DepthMaxF;
+ }
+#endif
+
+ return ~mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0];
+}
+
+
diff --git a/src/gallium/drivers/cell/spu/spu_exec.h b/src/gallium/drivers/cell/spu/spu_exec.h
new file mode 100644
index 00000000000..86056799405
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_exec.h
@@ -0,0 +1,172 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+#if !defined SPU_EXEC_H
+#define SPU_EXEC_H
+
+#include "pipe/p_compiler.h"
+#include "tgsi/tgsi_exec.h"
+
+#if defined __cplusplus
+extern "C" {
+#endif
+
+/**
+ * Registers may be treated as float, signed int or unsigned int.
+ */
+union spu_exec_channel
+{
+ float f[QUAD_SIZE];
+ int i[QUAD_SIZE];
+ unsigned u[QUAD_SIZE];
+ qword q;
+};
+
+/**
+ * A vector[RGBA] of channels[4 pixels]
+ */
+struct spu_exec_vector
+{
+ union spu_exec_channel xyzw[NUM_CHANNELS];
+};
+
+/**
+ * For fragment programs, information for computing fragment input
+ * values from plane equation of the triangle/line.
+ */
+struct spu_interp_coef
+{
+ float a0[NUM_CHANNELS]; /* in an xyzw layout */
+ float dadx[NUM_CHANNELS];
+ float dady[NUM_CHANNELS];
+};
+
+
+struct softpipe_tile_cache; /**< Opaque to TGSI */
+
+/**
+ * Information for sampling textures, which must be implemented
+ * by code outside the TGSI executor.
+ */
+struct spu_sampler
+{
+ const struct pipe_sampler_state *state;
+ struct pipe_texture *texture;
+ /** Get samples for four fragments in a quad */
+ void (*get_samples)(struct spu_sampler *sampler,
+ const float s[QUAD_SIZE],
+ const float t[QUAD_SIZE],
+ const float p[QUAD_SIZE],
+ float lodbias,
+ float rgba[NUM_CHANNELS][QUAD_SIZE]);
+ void *pipe; /*XXX temporary*/
+ struct softpipe_tile_cache *cache;
+};
+
+
+/**
+ * Run-time virtual machine state for executing TGSI shader.
+ */
+struct spu_exec_machine
+{
+ /*
+ * 32 program temporaries
+ * 4 internal temporaries
+ * 1 address
+ */
+ struct spu_exec_vector Temps[TGSI_EXEC_NUM_TEMPS
+ + TGSI_EXEC_NUM_TEMP_EXTRAS + 1]
+ ALIGN16_ATTRIB;
+
+ struct spu_exec_vector *Addrs;
+
+ struct spu_sampler *Samplers;
+
+ float Imms[TGSI_EXEC_NUM_IMMEDIATES][4];
+ unsigned ImmLimit;
+ float (*Consts)[4];
+ struct spu_exec_vector *Inputs;
+ struct spu_exec_vector *Outputs;
+ unsigned Processor;
+
+ /* GEOMETRY processor only. */
+ unsigned *Primitives;
+
+ /* FRAGMENT processor only. */
+ const struct spu_interp_coef *InterpCoefs;
+ struct spu_exec_vector QuadPos;
+
+ /* Conditional execution masks */
+ uint CondMask; /**< For IF/ELSE/ENDIF */
+ uint LoopMask; /**< For BGNLOOP/ENDLOOP */
+ uint ContMask; /**< For loop CONT statements */
+ uint FuncMask; /**< For function calls */
+ uint ExecMask; /**< = CondMask & LoopMask */
+
+ /** Condition mask stack (for nested conditionals) */
+ uint CondStack[TGSI_EXEC_MAX_COND_NESTING];
+ int CondStackTop;
+
+ /** Loop mask stack (for nested loops) */
+ uint LoopStack[TGSI_EXEC_MAX_LOOP_NESTING];
+ int LoopStackTop;
+
+ /** Loop continue mask stack (see comments in tgsi_exec.c) */
+ uint ContStack[TGSI_EXEC_MAX_LOOP_NESTING];
+ int ContStackTop;
+
+ /** Function execution mask stack (for executing subroutine code) */
+ uint FuncStack[TGSI_EXEC_MAX_CALL_NESTING];
+ int FuncStackTop;
+
+ /** Function call stack for saving/restoring the program counter */
+ uint CallStack[TGSI_EXEC_MAX_CALL_NESTING];
+ int CallStackTop;
+
+ struct tgsi_full_instruction *Instructions;
+ uint NumInstructions;
+
+ struct tgsi_full_declaration *Declarations;
+ uint NumDeclarations;
+};
+
+
+extern void
+spu_exec_machine_init(struct spu_exec_machine *mach,
+ uint numSamplers,
+ struct spu_sampler *samplers,
+ unsigned processor);
+
+extern uint
+spu_exec_machine_run( struct spu_exec_machine *mach );
+
+
+#if defined __cplusplus
+} /* extern "C" */
+#endif
+
+#endif /* SPU_EXEC_H */
diff --git a/src/gallium/drivers/cell/spu/spu_funcs.c b/src/gallium/drivers/cell/spu/spu_funcs.c
new file mode 100644
index 00000000000..ff3d609d258
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_funcs.c
@@ -0,0 +1,173 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+
+/**
+ * SPU functions accessed by shaders.
+ *
+ * Authors: Brian Paul
+ */
+
+
+#include <string.h>
+#include <libmisc.h>
+#include <math.h>
+#include <cos14_v.h>
+#include <sin14_v.h>
+#include <simdmath/exp2f4.h>
+#include <simdmath/log2f4.h>
+#include <simdmath/powf4.h>
+
+#include "cell/common.h"
+#include "spu_main.h"
+#include "spu_funcs.h"
+#include "spu_texture.h"
+
+
+/** For "return"-ing four vectors */
+struct vec_4x4
+{
+ vector float v[4];
+};
+
+
+static vector float
+spu_cos(vector float x)
+{
+ return _cos14_v(x);
+}
+
+static vector float
+spu_sin(vector float x)
+{
+ return _sin14_v(x);
+}
+
+static vector float
+spu_pow(vector float x, vector float y)
+{
+ return _powf4(x, y);
+}
+
+static vector float
+spu_exp2(vector float x)
+{
+ return _exp2f4(x);
+}
+
+static vector float
+spu_log2(vector float x)
+{
+ return _log2f4(x);
+}
+
+
+static struct vec_4x4
+spu_tex_2d(vector float s, vector float t, vector float r, vector float q,
+ unsigned unit)
+{
+ struct vec_4x4 colors;
+ (void) r;
+ (void) q;
+ spu.sample_texture_2d[unit](s, t, unit, 0, 0, colors.v);
+ return colors;
+}
+
+static struct vec_4x4
+spu_tex_3d(vector float s, vector float t, vector float r, vector float q,
+ unsigned unit)
+{
+ struct vec_4x4 colors;
+ (void) r;
+ (void) q;
+ spu.sample_texture_2d[unit](s, t, unit, 0, 0, colors.v);
+ return colors;
+}
+
+static struct vec_4x4
+spu_tex_cube(vector float s, vector float t, vector float r, vector float q,
+ unsigned unit)
+{
+ struct vec_4x4 colors;
+ (void) q;
+ sample_texture_cube(s, t, r, unit, colors.v);
+ return colors;
+}
+
+
+/**
+ * Add named function to list of "exported" functions that will be
+ * made available to the PPU-hosted code generator.
+ */
+static void
+export_func(struct cell_spu_function_info *spu_functions,
+ const char *name, void *addr)
+{
+ uint n = spu_functions->num;
+ ASSERT(strlen(name) < 16);
+ strcpy(spu_functions->names[n], name);
+ spu_functions->addrs[n] = (uint) addr;
+ spu_functions->num++;
+ ASSERT(spu_functions->num <= 16);
+}
+
+
+/**
+ * Return info about the SPU's function to the PPU / main memory.
+ * The PPU needs to know the address of some SPU-side functions so
+ * that we can generate shader code with function calls.
+ */
+void
+return_function_info(void)
+{
+ struct cell_spu_function_info funcs ALIGN16_ATTRIB;
+ int tag = TAG_MISC;
+
+ ASSERT(sizeof(funcs) == 256); /* must be multiple of 16 bytes */
+
+ funcs.num = 0;
+ export_func(&funcs, "spu_cos", &spu_cos);
+ export_func(&funcs, "spu_sin", &spu_sin);
+ export_func(&funcs, "spu_pow", &spu_pow);
+ export_func(&funcs, "spu_exp2", &spu_exp2);
+ export_func(&funcs, "spu_log2", &spu_log2);
+ export_func(&funcs, "spu_tex_2d", &spu_tex_2d);
+ export_func(&funcs, "spu_tex_3d", &spu_tex_3d);
+ export_func(&funcs, "spu_tex_cube", &spu_tex_cube);
+
+ /* Send the function info back to the PPU / main memory */
+ mfc_put((void *) &funcs, /* src in local store */
+ (unsigned int) spu.init.spu_functions, /* dst in main memory */
+ sizeof(funcs), /* bytes */
+ tag,
+ 0, /* tid */
+ 0 /* rid */);
+ wait_on_mask(1 << tag);
+}
+
+
+
diff --git a/src/gallium/drivers/cell/spu/spu_funcs.h b/src/gallium/drivers/cell/spu/spu_funcs.h
new file mode 100644
index 00000000000..3adb6ae99f9
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_funcs.h
@@ -0,0 +1,35 @@
+/**************************************************************************
+ *
+ * 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 SPU_FUNCS_H
+#define SPU_FUNCS_H
+
+extern void
+return_function_info(void);
+
+#endif
+
diff --git a/src/gallium/drivers/cell/spu/spu_main.c b/src/gallium/drivers/cell/spu/spu_main.c
new file mode 100644
index 00000000000..97c86d194da
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_main.c
@@ -0,0 +1,117 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+
+/* main() for Cell SPU code */
+
+
+#include <stdio.h>
+#include <libmisc.h>
+
+#include "pipe/p_defines.h"
+
+#include "spu_funcs.h"
+#include "spu_command.h"
+#include "spu_main.h"
+#include "spu_per_fragment_op.h"
+#include "spu_texture.h"
+//#include "spu_test.h"
+#include "cell/common.h"
+
+
+/*
+helpful headers:
+/usr/lib/gcc/spu/4.1.1/include/spu_mfcio.h
+/opt/cell/sdk/usr/include/libmisc.h
+*/
+
+struct spu_global spu;
+
+
+static void
+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();
+}
+
+/* In some versions of the SDK the SPE main takes 'unsigned long' as a
+ * parameter. In others it takes 'unsigned long long'. Use a define to
+ * select between the two.
+ */
+#ifdef SPU_MAIN_PARAM_LONG_LONG
+typedef unsigned long long main_param_t;
+#else
+typedef unsigned long main_param_t;
+#endif
+
+/**
+ * SPE entrypoint.
+ */
+int
+main(main_param_t speid, main_param_t argp)
+{
+ int tag = 0;
+
+ (void) speid;
+
+ ASSERT(sizeof(tile_t) == TILE_SIZE * TILE_SIZE * 4);
+ ASSERT(sizeof(struct cell_command_render) % 8 == 0);
+ ASSERT(sizeof(struct cell_command_fragment_ops) % 8 == 0);
+ ASSERT(((unsigned long) &spu.fragment_program_code) % 8 == 0);
+
+ one_time_init();
+ spu_command_init();
+
+ D_PRINTF(CELL_DEBUG_CMD, "main() speid=%lu\n", (unsigned long) speid);
+ D_PRINTF(CELL_DEBUG_FRAGMENT_OP_FALLBACK, "using fragment op fallback\n");
+
+ /* get initialization data */
+ mfc_get(&spu.init, /* dest */
+ (unsigned int) argp, /* src */
+ sizeof(struct cell_init_info), /* bytes */
+ tag,
+ 0, /* tid */
+ 0 /* rid */);
+ wait_on_mask( 1 << tag );
+
+ if (spu.init.id == 0) {
+ return_function_info();
+ }
+
+#if 0
+ if (spu.init.id==0)
+ spu_test_misc(spu.init.id);
+#endif
+
+ command_loop();
+
+ spu_command_close();
+
+ return 0;
+}
diff --git a/src/gallium/drivers/cell/spu/spu_main.h b/src/gallium/drivers/cell/spu/spu_main.h
new file mode 100644
index 00000000000..33767e7c51d
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_main.h
@@ -0,0 +1,254 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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 SPU_MAIN_H
+#define SPU_MAIN_H
+
+
+#include <spu_mfcio.h>
+
+#include "cell/common.h"
+#include "draw/draw_vertex.h"
+#include "pipe/p_state.h"
+
+
+#if DEBUG
+/* These debug macros use the unusual construction ", ##__VA_ARGS__"
+ * which expands to the expected comma + args if variadic arguments
+ * are supplied, but swallows the comma if there are no variadic
+ * arguments (which avoids syntax errors that would otherwise occur).
+ */
+#define D_PRINTF(flag, format,...) \
+ if (spu.init.debug_flags & (flag)) \
+ printf("SPU %u: " format, spu.init.id, ##__VA_ARGS__)
+#else
+#define D_PRINTF(...)
+#endif
+
+
+/**
+ * A tile is basically a TILE_SIZE x TILE_SIZE block of 4-byte pixels.
+ * The data may be addressed through several different types.
+ */
+typedef union {
+ ushort us[TILE_SIZE][TILE_SIZE];
+ uint ui[TILE_SIZE][TILE_SIZE];
+ vector unsigned short us8[TILE_SIZE/2][TILE_SIZE/4];
+ vector unsigned int ui4[TILE_SIZE/2][TILE_SIZE/2];
+} tile_t;
+
+
+#define TILE_STATUS_CLEAR 1
+#define TILE_STATUS_DEFINED 2 /**< defined in FB, but not in local store */
+#define TILE_STATUS_CLEAN 3 /**< in local store, but not changed */
+#define TILE_STATUS_DIRTY 4 /**< modified locally, but not put back yet */
+#define TILE_STATUS_GETTING 5 /**< mfc_get() called but not yet arrived */
+
+
+/** Function for sampling textures */
+typedef void (*spu_sample_texture_2d_func)(vector float s,
+ vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4]);
+
+
+/** Function for performing per-fragment ops */
+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);
+
+/** Function for running fragment program */
+typedef vector unsigned int (*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 */
+ void *depth_start; /**< addr of depth surface in main memory */
+ enum pipe_format color_format;
+ enum pipe_format depth_format;
+ uint width, height; /**< size in pixels */
+ uint width_tiles, height_tiles; /**< width and height in tiles */
+
+ uint color_clear_value;
+ uint depth_clear_value;
+
+ uint zsize; /**< 0, 2 or 4 bytes per Z */
+ float zscale; /**< 65535.0, 2^24-1 or 2^32-1 */
+} ALIGN16_ATTRIB;
+
+
+/** per-texture level info */
+struct spu_texture_level
+{
+ void *start;
+ ushort width, height, depth;
+ ushort tiles_per_row;
+ uint bytes_per_image;
+ /** texcoord scale factors */
+ vector float scale_s, scale_t, scale_r;
+ /** texcoord masks (if REPEAT then size-1, else ~0) */
+ vector signed int mask_s, mask_t, mask_r;
+ /** texcoord clamp limits */
+ vector signed int max_s, max_t, max_r;
+} ALIGN16_ATTRIB;
+
+
+struct spu_texture
+{
+ struct spu_texture_level level[CELL_MAX_TEXTURE_LEVELS];
+ uint max_level;
+ uint target; /**< PIPE_TEXTURE_x */
+} ALIGN16_ATTRIB;
+
+
+/**
+ * All SPU global/context state will be in a singleton object of this type:
+ */
+struct spu_global
+{
+ /** One-time init/constant info */
+ struct cell_init_info init;
+
+ /*
+ * Current state
+ */
+ struct spu_framebuffer fb;
+ struct pipe_depth_stencil_alpha_state depth_stencil_alpha;
+ struct pipe_blend_state blend;
+ struct pipe_blend_color blend_color;
+ struct pipe_sampler_state sampler[PIPE_MAX_SAMPLERS];
+ struct pipe_rasterizer_state rasterizer;
+ struct spu_texture texture[PIPE_MAX_SAMPLERS];
+ struct vertex_info vertex_info;
+
+ /** Current color and Z tiles */
+ tile_t ctile ALIGN16_ATTRIB;
+ tile_t ztile ALIGN16_ATTRIB;
+
+ /** Read depth/stencil tiles? */
+ boolean read_depth_stencil;
+
+ /** Current tiles' status */
+ ubyte cur_ctile_status, cur_ztile_status;
+
+ /** Status of all tiles in framebuffer */
+ ubyte ctile_status[CELL_MAX_HEIGHT/TILE_SIZE][CELL_MAX_WIDTH/TILE_SIZE] ALIGN16_ATTRIB;
+ 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;
+ uint fragment_ops_code_size;
+ /** 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;
+ /** Current fragment ops function */
+ spu_fragment_program_func fragment_program;
+
+ /** Current texture sampler function */
+ spu_sample_texture_2d_func sample_texture_2d[CELL_MAX_SAMPLERS];
+ spu_sample_texture_2d_func min_sample_texture_2d[CELL_MAX_SAMPLERS];
+ spu_sample_texture_2d_func mag_sample_texture_2d[CELL_MAX_SAMPLERS];
+
+ /** Fragment program constants */
+ vector float constants[4 * CELL_MAX_CONSTANTS];
+
+} ALIGN16_ATTRIB;
+
+
+extern struct spu_global spu;
+
+
+
+/* DMA TAGS */
+
+#define TAG_SURFACE_CLEAR 10
+#define TAG_VERTEX_BUFFER 11
+#define TAG_READ_TILE_COLOR 12
+#define TAG_READ_TILE_Z 13
+#define TAG_WRITE_TILE_COLOR 14
+#define TAG_WRITE_TILE_Z 15
+#define TAG_INDEX_BUFFER 16
+#define TAG_BATCH_BUFFER 17
+#define TAG_MISC 18
+#define TAG_DCACHE0 20
+#define TAG_DCACHE1 21
+#define TAG_DCACHE2 22
+#define TAG_DCACHE3 23
+#define TAG_FENCE 24
+
+
+static INLINE void
+wait_on_mask(unsigned tagMask)
+{
+ mfc_write_tag_mask( tagMask );
+ /* wait for completion of _any_ DMAs specified by tagMask */
+ mfc_read_tag_status_any();
+}
+
+
+static INLINE void
+wait_on_mask_all(unsigned tagMask)
+{
+ mfc_write_tag_mask( tagMask );
+ /* wait for completion of _any_ DMAs specified by tagMask */
+ mfc_read_tag_status_all();
+}
+
+
+
+
+
+static INLINE void
+memset16(ushort *d, ushort value, uint count)
+{
+ uint i;
+ for (i = 0; i < count; i++)
+ d[i] = value;
+}
+
+
+static INLINE void
+memset32(uint *d, uint value, uint count)
+{
+ uint i;
+ for (i = 0; i < count; i++)
+ d[i] = value;
+}
+
+
+#endif /* SPU_MAIN_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
new file mode 100644
index 00000000000..eba9f95cf1f
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.c
@@ -0,0 +1,631 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+/**
+ * \author Brian Paul
+ */
+
+
+#include <transpose_matrix4x4.h>
+#include "pipe/p_format.h"
+#include "spu_main.h"
+#include "spu_colorpack.h"
+#include "spu_per_fragment_op.h"
+
+
+#define LINEAR_QUAD_LAYOUT 1
+
+
+static INLINE vector float
+spu_min(vector float a, vector float b)
+{
+ vector unsigned int m;
+ m = spu_cmpgt(a, b); /* m = a > b ? ~0 : 0 */
+ return spu_sel(a, b, m);
+}
+
+
+static INLINE vector float
+spu_max(vector float a, vector float b)
+{
+ vector unsigned int m;
+ m = spu_cmpgt(a, b); /* m = a > b ? ~0 : 0 */
+ return spu_sel(b, a, m);
+}
+
+
+/**
+ * Called by rasterizer for each quad after the shader has run. Do
+ * all the per-fragment operations including alpha test, z test,
+ * stencil test, blend, colormask and logicops. 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 fragR,
+ vector float fragG,
+ vector float fragB,
+ vector float fragA,
+ vector unsigned int mask)
+{
+ vector float frag_aos[4];
+ unsigned int fbc0, fbc1, fbc2, fbc3 ; /* framebuffer/tile colors */
+ unsigned int fragc0, fragc1, fragc2, fragc3; /* fragment colors */
+
+ /*
+ * Do alpha test
+ */
+ if (spu.depth_stencil_alpha.alpha.enabled) {
+ vector float ref = spu_splats(spu.depth_stencil_alpha.alpha.ref_value);
+ vector unsigned int amask;
+
+ switch (spu.depth_stencil_alpha.alpha.func) {
+ case PIPE_FUNC_LESS:
+ amask = spu_cmpgt(ref, fragA); /* mask = (fragA < ref) */
+ break;
+ case PIPE_FUNC_GREATER:
+ amask = spu_cmpgt(fragA, ref); /* mask = (fragA > ref) */
+ break;
+ case PIPE_FUNC_GEQUAL:
+ amask = spu_cmpgt(ref, fragA);
+ amask = spu_nor(amask, amask);
+ break;
+ case PIPE_FUNC_LEQUAL:
+ amask = spu_cmpgt(fragA, ref);
+ amask = spu_nor(amask, amask);
+ break;
+ case PIPE_FUNC_EQUAL:
+ amask = spu_cmpeq(ref, fragA);
+ break;
+ case PIPE_FUNC_NOTEQUAL:
+ amask = spu_cmpeq(ref, fragA);
+ 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;
+ }
+ }
+
+
+ /*
+ * If we'll need the current framebuffer/tile colors for blending
+ * or logicop or colormask, fetch them now.
+ */
+ if (spu.blend.blend_enable ||
+ spu.blend.logicop_enable ||
+ spu.blend.colormask != 0xf) {
+
+#if LINEAR_QUAD_LAYOUT /* See comments/diagram below */
+ fbc0 = colorTile->ui[y][x*2+0];
+ fbc1 = colorTile->ui[y][x*2+1];
+ fbc2 = colorTile->ui[y][x*2+2];
+ fbc3 = colorTile->ui[y][x*2+3];
+#else
+ fbc0 = colorTile->ui[y+0][x+0];
+ fbc1 = colorTile->ui[y+0][x+1];
+ fbc2 = colorTile->ui[y+1][x+0];
+ fbc3 = colorTile->ui[y+1][x+1];
+#endif
+ }
+
+
+ /*
+ * Do blending
+ */
+ if (spu.blend.blend_enable) {
+ /* blending terms, misc regs */
+ vector float term1r, term1g, term1b, term1a;
+ vector float term2r, term2g, term2b, term2a;
+ vector float one, tmp;
+
+ vector float fbRGBA[4]; /* current framebuffer colors */
+
+ /* convert framebuffer colors from packed int to vector float */
+ {
+ vector float temp[4]; /* float colors in AOS form */
+ switch (spu.fb.color_format) {
+ case PIPE_FORMAT_B8G8R8A8_UNORM:
+ temp[0] = spu_unpack_B8G8R8A8(fbc0);
+ temp[1] = spu_unpack_B8G8R8A8(fbc1);
+ temp[2] = spu_unpack_B8G8R8A8(fbc2);
+ temp[3] = spu_unpack_B8G8R8A8(fbc3);
+ break;
+ case PIPE_FORMAT_A8R8G8B8_UNORM:
+ temp[0] = spu_unpack_A8R8G8B8(fbc0);
+ temp[1] = spu_unpack_A8R8G8B8(fbc1);
+ temp[2] = spu_unpack_A8R8G8B8(fbc2);
+ temp[3] = spu_unpack_A8R8G8B8(fbc3);
+ break;
+ default:
+ ASSERT(0);
+ }
+ _transpose_matrix4x4(fbRGBA, temp); /* fbRGBA = transpose(temp) */
+ }
+
+ /*
+ * Compute Src RGB terms (fragment color * factor)
+ */
+ switch (spu.blend.rgb_src_factor) {
+ case PIPE_BLENDFACTOR_ONE:
+ term1r = fragR;
+ term1g = fragG;
+ term1b = fragB;
+ break;
+ case PIPE_BLENDFACTOR_ZERO:
+ term1r =
+ term1g =
+ term1b = spu_splats(0.0f);
+ break;
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ term1r = spu_mul(fragR, fragR);
+ term1g = spu_mul(fragG, fragG);
+ term1b = spu_mul(fragB, fragB);
+ break;
+ case PIPE_BLENDFACTOR_SRC_ALPHA:
+ term1r = spu_mul(fragR, fragA);
+ term1g = spu_mul(fragG, fragA);
+ term1b = spu_mul(fragB, fragA);
+ break;
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ term1r = spu_mul(fragR, fbRGBA[0]);
+ term1g = spu_mul(fragG, fbRGBA[1]);
+ term1b = spu_mul(fragB, fbRGBA[1]);
+ break;
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ term1r = spu_mul(fragR, fbRGBA[3]);
+ term1g = spu_mul(fragG, fbRGBA[3]);
+ term1b = spu_mul(fragB, fbRGBA[3]);
+ break;
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ term1r = spu_mul(fragR, spu_splats(spu.blend_color.color[0]));
+ term1g = spu_mul(fragG, spu_splats(spu.blend_color.color[1]));
+ term1b = spu_mul(fragB, spu_splats(spu.blend_color.color[2]));
+ break;
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ term1r = spu_mul(fragR, spu_splats(spu.blend_color.color[3]));
+ term1g = spu_mul(fragG, spu_splats(spu.blend_color.color[3]));
+ term1b = spu_mul(fragB, spu_splats(spu.blend_color.color[3]));
+ break;
+ /* XXX more cases */
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Compute Src Alpha term (fragment alpha * factor)
+ */
+ switch (spu.blend.alpha_src_factor) {
+ case PIPE_BLENDFACTOR_ONE:
+ term1a = fragA;
+ break;
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ term1a = spu_splats(0.0f);
+ break;
+ case PIPE_BLENDFACTOR_SRC_ALPHA:
+ term1a = spu_mul(fragA, fragA);
+ break;
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ /* fall-through */
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ term1a = spu_mul(fragA, fbRGBA[3]);
+ break;
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ /* fall-through */
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ term1a = spu_mul(fragR, spu_splats(spu.blend_color.color[3]));
+ break;
+ /* XXX more cases */
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Compute Dest RGB terms (framebuffer color * factor)
+ */
+ switch (spu.blend.rgb_dst_factor) {
+ case PIPE_BLENDFACTOR_ONE:
+ term2r = fbRGBA[0];
+ term2g = fbRGBA[1];
+ term2b = fbRGBA[2];
+ break;
+ case PIPE_BLENDFACTOR_ZERO:
+ term2r =
+ term2g =
+ term2b = spu_splats(0.0f);
+ break;
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ term2r = spu_mul(fbRGBA[0], fragR);
+ term2g = spu_mul(fbRGBA[1], fragG);
+ term2b = spu_mul(fbRGBA[2], fragB);
+ break;
+ case PIPE_BLENDFACTOR_SRC_ALPHA:
+ term2r = spu_mul(fbRGBA[0], fragA);
+ term2g = spu_mul(fbRGBA[1], fragA);
+ term2b = spu_mul(fbRGBA[2], fragA);
+ break;
+ case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
+ one = spu_splats(1.0f);
+ tmp = spu_sub(one, fragA);
+ term2r = spu_mul(fbRGBA[0], tmp);
+ term2g = spu_mul(fbRGBA[1], tmp);
+ term2b = spu_mul(fbRGBA[2], tmp);
+ break;
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ term2r = spu_mul(fbRGBA[0], fbRGBA[0]);
+ term2g = spu_mul(fbRGBA[1], fbRGBA[1]);
+ term2b = spu_mul(fbRGBA[2], fbRGBA[2]);
+ break;
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ term2r = spu_mul(fbRGBA[0], fbRGBA[3]);
+ term2g = spu_mul(fbRGBA[1], fbRGBA[3]);
+ term2b = spu_mul(fbRGBA[2], fbRGBA[3]);
+ break;
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ term2r = spu_mul(fbRGBA[0], spu_splats(spu.blend_color.color[0]));
+ term2g = spu_mul(fbRGBA[1], spu_splats(spu.blend_color.color[1]));
+ term2b = spu_mul(fbRGBA[2], spu_splats(spu.blend_color.color[2]));
+ break;
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ term2r = spu_mul(fbRGBA[0], spu_splats(spu.blend_color.color[3]));
+ term2g = spu_mul(fbRGBA[1], spu_splats(spu.blend_color.color[3]));
+ term2b = spu_mul(fbRGBA[2], spu_splats(spu.blend_color.color[3]));
+ break;
+ /* XXX more cases */
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Compute Dest Alpha term (framebuffer alpha * factor)
+ */
+ switch (spu.blend.alpha_dst_factor) {
+ case PIPE_BLENDFACTOR_ONE:
+ term2a = fbRGBA[3];
+ break;
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ term2a = spu_splats(0.0f);
+ break;
+ case PIPE_BLENDFACTOR_SRC_ALPHA:
+ term2a = spu_mul(fbRGBA[3], fragA);
+ break;
+ case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
+ one = spu_splats(1.0f);
+ tmp = spu_sub(one, fragA);
+ term2a = spu_mul(fbRGBA[3], tmp);
+ break;
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ /* fall-through */
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ term2a = spu_mul(fbRGBA[3], fbRGBA[3]);
+ break;
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ /* fall-through */
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ term2a = spu_mul(fbRGBA[3], spu_splats(spu.blend_color.color[3]));
+ break;
+ /* XXX more cases */
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Combine Src/Dest RGB terms
+ */
+ switch (spu.blend.rgb_func) {
+ case PIPE_BLEND_ADD:
+ fragR = spu_add(term1r, term2r);
+ fragG = spu_add(term1g, term2g);
+ fragB = spu_add(term1b, term2b);
+ break;
+ case PIPE_BLEND_SUBTRACT:
+ fragR = spu_sub(term1r, term2r);
+ fragG = spu_sub(term1g, term2g);
+ fragB = spu_sub(term1b, term2b);
+ break;
+ case PIPE_BLEND_REVERSE_SUBTRACT:
+ fragR = spu_sub(term2r, term1r);
+ fragG = spu_sub(term2g, term1g);
+ fragB = spu_sub(term2b, term1b);
+ break;
+ case PIPE_BLEND_MIN:
+ fragR = spu_min(term1r, term2r);
+ fragG = spu_min(term1g, term2g);
+ fragB = spu_min(term1b, term2b);
+ break;
+ case PIPE_BLEND_MAX:
+ fragR = spu_max(term1r, term2r);
+ fragG = spu_max(term1g, term2g);
+ fragB = spu_max(term1b, term2b);
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Combine Src/Dest A term
+ */
+ switch (spu.blend.alpha_func) {
+ case PIPE_BLEND_ADD:
+ fragA = spu_add(term1a, term2a);
+ break;
+ case PIPE_BLEND_SUBTRACT:
+ fragA = spu_sub(term1a, term2a);
+ break;
+ case PIPE_BLEND_REVERSE_SUBTRACT:
+ fragA = spu_sub(term2a, term1a);
+ break;
+ case PIPE_BLEND_MIN:
+ fragA = spu_min(term1a, term2a);
+ break;
+ case PIPE_BLEND_MAX:
+ fragA = spu_max(term1a, term2a);
+ break;
+ default:
+ ASSERT(0);
+ }
+ }
+
+
+ /*
+ * Convert RRRR,GGGG,BBBB,AAAA to RGBA,RGBA,RGBA,RGBA.
+ */
+#if 0
+ /* original code */
+ {
+ vector float frag_soa[4];
+ frag_soa[0] = fragR;
+ frag_soa[1] = fragG;
+ frag_soa[2] = fragB;
+ frag_soa[3] = fragA;
+ _transpose_matrix4x4(frag_aos, frag_soa);
+ }
+#else
+ /* short-cut relying on function parameter layout: */
+ _transpose_matrix4x4(frag_aos, &fragR);
+ (void) fragG;
+ (void) fragB;
+#endif
+
+ /*
+ * Pack fragment float colors into 32-bit RGBA words.
+ */
+ switch (spu.fb.color_format) {
+ case PIPE_FORMAT_A8R8G8B8_UNORM:
+ fragc0 = spu_pack_A8R8G8B8(frag_aos[0]);
+ fragc1 = spu_pack_A8R8G8B8(frag_aos[1]);
+ fragc2 = spu_pack_A8R8G8B8(frag_aos[2]);
+ fragc3 = spu_pack_A8R8G8B8(frag_aos[3]);
+ break;
+ case PIPE_FORMAT_B8G8R8A8_UNORM:
+ fragc0 = spu_pack_B8G8R8A8(frag_aos[0]);
+ fragc1 = spu_pack_B8G8R8A8(frag_aos[1]);
+ fragc2 = spu_pack_B8G8R8A8(frag_aos[2]);
+ fragc3 = spu_pack_B8G8R8A8(frag_aos[3]);
+ break;
+ default:
+ fprintf(stderr, "SPU: Bad pixel format in spu_default_fragment_ops\n");
+ ASSERT(0);
+ }
+
+
+ /*
+ * Do color masking
+ */
+ if (spu.blend.colormask != 0xf) {
+ uint cmask = 0x0; /* each byte corresponds to a color channel */
+
+ /* Form bitmask depending on color buffer format and colormask bits */
+ switch (spu.fb.color_format) {
+ case PIPE_FORMAT_A8R8G8B8_UNORM:
+ if (spu.blend.colormask & PIPE_MASK_R)
+ cmask |= 0x00ff0000; /* red */
+ if (spu.blend.colormask & PIPE_MASK_G)
+ cmask |= 0x0000ff00; /* green */
+ if (spu.blend.colormask & PIPE_MASK_B)
+ cmask |= 0x000000ff; /* blue */
+ if (spu.blend.colormask & PIPE_MASK_A)
+ cmask |= 0xff000000; /* alpha */
+ break;
+ case PIPE_FORMAT_B8G8R8A8_UNORM:
+ if (spu.blend.colormask & PIPE_MASK_R)
+ cmask |= 0x0000ff00; /* red */
+ if (spu.blend.colormask & PIPE_MASK_G)
+ cmask |= 0x00ff0000; /* green */
+ if (spu.blend.colormask & PIPE_MASK_B)
+ cmask |= 0xff000000; /* blue */
+ if (spu.blend.colormask & PIPE_MASK_A)
+ cmask |= 0x000000ff; /* alpha */
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ /*
+ * Apply color mask to the 32-bit packed colors.
+ * if (cmask[i])
+ * frag color[i] = frag color[i];
+ * else
+ * frag color[i] = framebuffer color[i];
+ */
+ fragc0 = (fragc0 & cmask) | (fbc0 & ~cmask);
+ fragc1 = (fragc1 & cmask) | (fbc1 & ~cmask);
+ fragc2 = (fragc2 & cmask) | (fbc2 & ~cmask);
+ fragc3 = (fragc3 & cmask) | (fbc3 & ~cmask);
+ }
+
+
+ /*
+ * Do logic ops
+ */
+ if (spu.blend.logicop_enable) {
+ /* XXX to do */
+ /* apply logicop to 32-bit packed colors (fragcx and fbcx) */
+ }
+
+
+ /*
+ * If mask is non-zero, mark tile as dirty.
+ */
+ if (spu_extract(spu_orx(mask), 0)) {
+ spu.cur_ctile_status = TILE_STATUS_DIRTY;
+ }
+ else {
+ /* write no fragments */
+ return;
+ }
+
+
+ /*
+ * Write new fragment/quad colors to the framebuffer/tile.
+ * Only write pixels where the corresponding mask word is set.
+ */
+#if LINEAR_QUAD_LAYOUT
+ /*
+ * Quad layout:
+ * +--+--+--+--+
+ * |p0|p1|p2|p3|...
+ * +--+--+--+--+
+ */
+ if (spu_extract(mask, 0))
+ colorTile->ui[y][x*2] = fragc0;
+ if (spu_extract(mask, 1))
+ colorTile->ui[y][x*2+1] = fragc1;
+ if (spu_extract(mask, 2))
+ colorTile->ui[y][x*2+2] = fragc2;
+ if (spu_extract(mask, 3))
+ colorTile->ui[y][x*2+3] = fragc3;
+#else
+ /*
+ * Quad layout:
+ * +--+--+
+ * |p0|p1|...
+ * +--+--+
+ * |p2|p3|...
+ * +--+--+
+ */
+ if (spu_extract(mask, 0))
+ colorTile->ui[y+0][x+0] = fragc0;
+ if (spu_extract(mask, 1))
+ colorTile->ui[y+0][x+1] = fragc1;
+ if (spu_extract(mask, 2))
+ colorTile->ui[y+1][x+0] = fragc2;
+ if (spu_extract(mask, 3))
+ colorTile->ui[y+1][x+1] = fragc3;
+#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
new file mode 100644
index 00000000000..f817abf0463
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_per_fragment_op.h
@@ -0,0 +1,44 @@
+/**************************************************************************
+ *
+ * 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 SPU_PER_FRAGMENT_OP
+#define SPU_PER_FRAGMENT_OP
+
+
+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_render.c b/src/gallium/drivers/cell/spu/spu_render.c
new file mode 100644
index 00000000000..7c225e2f27c
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_render.c
@@ -0,0 +1,295 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+
+#include <stdio.h>
+#include <libmisc.h>
+#include <spu_mfcio.h>
+
+#include "spu_main.h"
+#include "spu_render.h"
+#include "spu_tri.h"
+#include "spu_tile.h"
+#include "cell/common.h"
+#include "util/u_memory.h"
+
+
+/**
+ * Given a rendering command's bounding box (in pixels) compute the
+ * location of the corresponding screen tile bounding box.
+ */
+static INLINE void
+tile_bounding_box(const struct cell_command_render *render,
+ uint *txmin, uint *tymin,
+ uint *box_num_tiles, uint *box_width_tiles)
+{
+#if 0
+ /* Debug: full-window bounding box */
+ uint txmax = spu.fb.width_tiles - 1;
+ uint tymax = spu.fb.height_tiles - 1;
+ *txmin = 0;
+ *tymin = 0;
+ *box_num_tiles = spu.fb.width_tiles * spu.fb.height_tiles;
+ *box_width_tiles = spu.fb.width_tiles;
+ (void) render;
+ (void) txmax;
+ (void) tymax;
+#else
+ uint txmax, tymax, box_height_tiles;
+
+ *txmin = (uint) render->xmin / TILE_SIZE;
+ *tymin = (uint) render->ymin / TILE_SIZE;
+ txmax = (uint) render->xmax / TILE_SIZE;
+ tymax = (uint) render->ymax / TILE_SIZE;
+ if (txmax >= spu.fb.width_tiles)
+ txmax = spu.fb.width_tiles-1;
+ if (tymax >= spu.fb.height_tiles)
+ tymax = spu.fb.height_tiles-1;
+ *box_width_tiles = txmax - *txmin + 1;
+ box_height_tiles = tymax - *tymin + 1;
+ *box_num_tiles = *box_width_tiles * box_height_tiles;
+#endif
+#if 0
+ printf("SPU %u: bounds: %g, %g ... %g, %g\n", spu.init.id,
+ render->xmin, render->ymin, render->xmax, render->ymax);
+ printf("SPU %u: tiles: %u, %u .. %u, %u\n",
+ spu.init.id, *txmin, *tymin, txmax, tymax);
+ ASSERT(render->xmin <= render->xmax);
+ ASSERT(render->ymin <= render->ymax);
+#endif
+}
+
+
+/** Check if the tile at (tx,ty) belongs to this SPU */
+static INLINE boolean
+my_tile(uint tx, uint ty)
+{
+ return (spu.fb.width_tiles * ty + tx) % spu.init.num_spus == spu.init.id;
+}
+
+
+/**
+ * Start fetching non-clear color/Z tiles from main memory
+ */
+static INLINE void
+get_cz_tiles(uint tx, uint ty)
+{
+ if (spu.read_depth_stencil) {
+ if (spu.cur_ztile_status != TILE_STATUS_CLEAR) {
+ //printf("SPU %u: getting Z tile %u, %u\n", spu.init.id, tx, ty);
+ get_tile(tx, ty, &spu.ztile, TAG_READ_TILE_Z, 1);
+ spu.cur_ztile_status = TILE_STATUS_GETTING;
+ }
+ }
+
+ if (spu.cur_ctile_status != TILE_STATUS_CLEAR) {
+ //printf("SPU %u: getting C tile %u, %u\n", spu.init.id, tx, ty);
+ get_tile(tx, ty, &spu.ctile, TAG_READ_TILE_COLOR, 0);
+ spu.cur_ctile_status = TILE_STATUS_GETTING;
+ }
+}
+
+
+/**
+ * Start putting dirty color/Z tiles back to main memory
+ */
+static INLINE void
+put_cz_tiles(uint tx, uint ty)
+{
+ if (spu.cur_ztile_status == TILE_STATUS_DIRTY) {
+ /* tile was modified and needs to be written back */
+ //printf("SPU %u: put dirty Z tile %u, %u\n", spu.init.id, tx, ty);
+ put_tile(tx, ty, &spu.ztile, TAG_WRITE_TILE_Z, 1);
+ spu.cur_ztile_status = TILE_STATUS_DEFINED;
+ }
+ else if (spu.cur_ztile_status == TILE_STATUS_GETTING) {
+ /* tile was never used */
+ spu.cur_ztile_status = TILE_STATUS_DEFINED;
+ //printf("SPU %u: put getting Z tile %u, %u\n", spu.init.id, tx, ty);
+ }
+
+ if (spu.cur_ctile_status == TILE_STATUS_DIRTY) {
+ /* tile was modified and needs to be written back */
+ //printf("SPU %u: put dirty C tile %u, %u\n", spu.init.id, tx, ty);
+ put_tile(tx, ty, &spu.ctile, TAG_WRITE_TILE_COLOR, 0);
+ spu.cur_ctile_status = TILE_STATUS_DEFINED;
+ }
+ else if (spu.cur_ctile_status == TILE_STATUS_GETTING) {
+ /* tile was never used */
+ spu.cur_ctile_status = TILE_STATUS_DEFINED;
+ //printf("SPU %u: put getting C tile %u, %u\n", spu.init.id, tx, ty);
+ }
+}
+
+
+/**
+ * Wait for 'put' of color/z tiles to complete.
+ */
+static INLINE void
+wait_put_cz_tiles(void)
+{
+ wait_on_mask(1 << TAG_WRITE_TILE_COLOR);
+ if (spu.read_depth_stencil) {
+ wait_on_mask(1 << TAG_WRITE_TILE_Z);
+ }
+}
+
+
+/**
+ * Render primitives
+ * \param pos_incr returns value indicating how may words to skip after
+ * this command in the batch buffer
+ */
+void
+cmd_render(const struct cell_command_render *render, uint *pos_incr)
+{
+ /* we'll DMA into these buffers */
+ ubyte vertex_data[CELL_BUFFER_SIZE] ALIGN16_ATTRIB;
+ const uint vertex_size = render->vertex_size; /* in bytes */
+ /*const*/ uint total_vertex_bytes = render->num_verts * vertex_size;
+ uint index_bytes;
+ const ubyte *vertices;
+ const ushort *indexes;
+ uint i, j;
+ uint num_tiles;
+
+ D_PRINTF(CELL_DEBUG_CMD,
+ "RENDER prim=%u num_vert=%u num_ind=%u inline_vert=%u\n",
+ render->prim_type,
+ render->num_verts,
+ render->num_indexes,
+ render->inline_verts);
+
+ ASSERT(sizeof(*render) % 4 == 0);
+ ASSERT(total_vertex_bytes % 16 == 0);
+ ASSERT(render->prim_type == PIPE_PRIM_TRIANGLES);
+ ASSERT(render->num_indexes % 3 == 0);
+
+
+ /* indexes are right after the render command in the batch buffer */
+ indexes = (const ushort *) (render + 1);
+ index_bytes = ROUNDUP8(render->num_indexes * 2);
+ *pos_incr = index_bytes / 8 + sizeof(*render) / 8;
+
+
+ if (render->inline_verts) {
+ /* Vertices are after indexes in batch buffer at next 16-byte addr */
+ vertices = (const ubyte *) render + (*pos_incr * 8);
+ vertices = (const ubyte *) align_pointer((void *) vertices, 16);
+ ASSERT_ALIGN16(vertices);
+ *pos_incr = ((vertices + total_vertex_bytes) - (ubyte *) render) / 8;
+ }
+ else {
+ /* Begin DMA fetch of vertex buffer */
+ ubyte *src = spu.init.buffers[render->vertex_buf];
+ ubyte *dest = vertex_data;
+
+ /* skip vertex data we won't use */
+#if 01
+ src += render->min_index * vertex_size;
+ dest += render->min_index * vertex_size;
+ total_vertex_bytes -= render->min_index * vertex_size;
+#endif
+ ASSERT(total_vertex_bytes % 16 == 0);
+ ASSERT_ALIGN16(dest);
+ ASSERT_ALIGN16(src);
+
+ mfc_get(dest, /* in vertex_data[] array */
+ (unsigned int) src, /* src in main memory */
+ total_vertex_bytes, /* size */
+ TAG_VERTEX_BUFFER,
+ 0, /* tid */
+ 0 /* rid */);
+
+ vertices = vertex_data;
+
+ wait_on_mask(1 << TAG_VERTEX_BUFFER);
+ }
+
+
+ /**
+ ** find tiles which intersect the prim bounding box
+ **/
+ uint txmin, tymin, box_width_tiles, box_num_tiles;
+ tile_bounding_box(render, &txmin, &tymin,
+ &box_num_tiles, &box_width_tiles);
+
+
+ /* make sure any pending clears have completed */
+ wait_on_mask(1 << TAG_SURFACE_CLEAR); /* XXX temporary */
+
+
+ num_tiles = 0;
+
+ /**
+ ** loop over tiles, rendering tris
+ **/
+ for (i = 0; i < box_num_tiles; i++) {
+ const uint tx = txmin + i % box_width_tiles;
+ const uint ty = tymin + i / box_width_tiles;
+
+ ASSERT(tx < spu.fb.width_tiles);
+ ASSERT(ty < spu.fb.height_tiles);
+
+ if (!my_tile(tx, ty))
+ continue;
+
+ num_tiles++;
+
+ spu.cur_ctile_status = spu.ctile_status[ty][tx];
+ spu.cur_ztile_status = spu.ztile_status[ty][tx];
+
+ get_cz_tiles(tx, ty);
+
+ uint drawn = 0;
+
+ /* loop over tris */
+ for (j = 0; j < render->num_indexes; j += 3) {
+ const float *v0, *v1, *v2;
+
+ v0 = (const float *) (vertices + indexes[j+0] * vertex_size);
+ 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);
+ }
+
+ //printf("SPU %u: drew %u of %u\n", spu.init.id, drawn, render->num_indexes/3);
+
+ /* write color/z tiles back to main framebuffer, if dirtied */
+ put_cz_tiles(tx, ty);
+
+ wait_put_cz_tiles(); /* XXX seems unnecessary... */
+
+ spu.ctile_status[ty][tx] = spu.cur_ctile_status;
+ spu.ztile_status[ty][tx] = spu.cur_ztile_status;
+ }
+
+ D_PRINTF(CELL_DEBUG_CMD,
+ "RENDER done (%u tiles hit)\n",
+ num_tiles);
+}
diff --git a/src/gallium/drivers/cell/spu/spu_render.h b/src/gallium/drivers/cell/spu/spu_render.h
new file mode 100644
index 00000000000..493434f0878
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_render.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 SPU_RENDER_H
+#define SPU_RENDER_H
+
+#include "cell/common.h"
+
+extern void
+cmd_render(const struct cell_command_render *render, uint *pos_incr);
+
+#endif /* SPU_RENDER_H */
+
diff --git a/src/gallium/drivers/cell/spu/spu_shuffle.h b/src/gallium/drivers/cell/spu/spu_shuffle.h
new file mode 100644
index 00000000000..74f2a0b6d2e
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_shuffle.h
@@ -0,0 +1,186 @@
+#ifndef SPU_SHUFFLE_H
+#define SPU_SHUFFLE_H
+
+/*
+ * Generate shuffle patterns with minimal fuss.
+ *
+ * Based on ideas from
+ * http://www.insomniacgames.com/tech/articles/0408/files/shuffles.pdf
+ *
+ * A-P indicates 0-15th position in first vector
+ * a-p indicates 0-15th position in second vector
+ *
+ * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+ * |00|01|02|03|04|05|06|07|08|09|0a|0b|0c|0d|0e|0f|
+ * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+ * | A| B| C| D|
+ * +-----+-----+-----+-----+-----+-----+-----+-----+
+ * | A| B| C| D| E| F| G| H|
+ * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+ * | A| B| C| D| E| F| G| H| I| J| K| L| M| N| O| P|
+ * +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+ *
+ * x or X indicates 0xff
+ * 8 indicates 0x80
+ * 0 indicates 0x00
+ *
+ * The macros SHUFFLE4() SHUFFLE8() and SHUFFLE16() provide a const vector
+ * unsigned char literal suitable for use with spu_shuffle().
+ *
+ * The macros SHUFB4() SHUFB8() and SHUFB16() provide a const qword vector
+ * literal suitable for use with si_shufb().
+ *
+ *
+ * For example :
+ * SHUFB4(A,A,A,A)
+ * expands to :
+ * ((const qword){0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3})
+ *
+ * SHUFFLE8(A,B,a,b,C,c,8,8)
+ * expands to :
+ * ((const vector unsigned char){0x00,0x01,0x02,0x03,0x10,0x11,0x12,0x13,
+ * 0x04,0x05,0x14,0x15,0xe0,0xe0,0xe0,0xe0})
+ *
+ */
+
+#include <spu_intrinsics.h>
+
+#define SHUFFLE_PATTERN_4_A__ 0x00, 0x01, 0x02, 0x03
+#define SHUFFLE_PATTERN_4_B__ 0x04, 0x05, 0x06, 0x07
+#define SHUFFLE_PATTERN_4_C__ 0x08, 0x09, 0x0a, 0x0b
+#define SHUFFLE_PATTERN_4_D__ 0x0c, 0x0d, 0x0e, 0x0f
+#define SHUFFLE_PATTERN_4_a__ 0x10, 0x11, 0x12, 0x13
+#define SHUFFLE_PATTERN_4_b__ 0x14, 0x15, 0x16, 0x17
+#define SHUFFLE_PATTERN_4_c__ 0x18, 0x19, 0x1a, 0x1b
+#define SHUFFLE_PATTERN_4_d__ 0x1c, 0x1d, 0x1e, 0x1f
+#define SHUFFLE_PATTERN_4_X__ 0xc0, 0xc0, 0xc0, 0xc0
+#define SHUFFLE_PATTERN_4_x__ 0xc0, 0xc0, 0xc0, 0xc0
+#define SHUFFLE_PATTERN_4_0__ 0x80, 0x80, 0x80, 0x80
+#define SHUFFLE_PATTERN_4_8__ 0xe0, 0xe0, 0xe0, 0xe0
+
+#define SHUFFLE_VECTOR_4__(A, B, C, D) \
+ SHUFFLE_PATTERN_4_##A##__, \
+ SHUFFLE_PATTERN_4_##B##__, \
+ SHUFFLE_PATTERN_4_##C##__, \
+ SHUFFLE_PATTERN_4_##D##__
+
+#define SHUFFLE4(A, B, C, D) \
+ ((const vector unsigned char){ \
+ SHUFFLE_VECTOR_4__(A, B, C, D) \
+ })
+
+#define SHUFB4(A, B, C, D) \
+ ((const qword){ \
+ SHUFFLE_VECTOR_4__(A, B, C, D) \
+ })
+
+
+#define SHUFFLE_PATTERN_8_A__ 0x00, 0x01
+#define SHUFFLE_PATTERN_8_B__ 0x02, 0x03
+#define SHUFFLE_PATTERN_8_C__ 0x04, 0x05
+#define SHUFFLE_PATTERN_8_D__ 0x06, 0x07
+#define SHUFFLE_PATTERN_8_E__ 0x08, 0x09
+#define SHUFFLE_PATTERN_8_F__ 0x0a, 0x0b
+#define SHUFFLE_PATTERN_8_G__ 0x0c, 0x0d
+#define SHUFFLE_PATTERN_8_H__ 0x0e, 0x0f
+#define SHUFFLE_PATTERN_8_a__ 0x10, 0x11
+#define SHUFFLE_PATTERN_8_b__ 0x12, 0x13
+#define SHUFFLE_PATTERN_8_c__ 0x14, 0x15
+#define SHUFFLE_PATTERN_8_d__ 0x16, 0x17
+#define SHUFFLE_PATTERN_8_e__ 0x18, 0x19
+#define SHUFFLE_PATTERN_8_f__ 0x1a, 0x1b
+#define SHUFFLE_PATTERN_8_g__ 0x1c, 0x1d
+#define SHUFFLE_PATTERN_8_h__ 0x1e, 0x1f
+#define SHUFFLE_PATTERN_8_X__ 0xc0, 0xc0
+#define SHUFFLE_PATTERN_8_x__ 0xc0, 0xc0
+#define SHUFFLE_PATTERN_8_0__ 0x80, 0x80
+#define SHUFFLE_PATTERN_8_8__ 0xe0, 0xe0
+
+
+#define SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \
+ SHUFFLE_PATTERN_8_##A##__, \
+ SHUFFLE_PATTERN_8_##B##__, \
+ SHUFFLE_PATTERN_8_##C##__, \
+ SHUFFLE_PATTERN_8_##D##__, \
+ SHUFFLE_PATTERN_8_##E##__, \
+ SHUFFLE_PATTERN_8_##F##__, \
+ SHUFFLE_PATTERN_8_##G##__, \
+ SHUFFLE_PATTERN_8_##H##__
+
+#define SHUFFLE8(A, B, C, D, E, F, G, H) \
+ ((const vector unsigned char){ \
+ SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \
+ })
+
+#define SHUFB8(A, B, C, D, E, F, G, H) \
+ ((const qword){ \
+ SHUFFLE_VECTOR_8__(A, B, C, D, E, F, G, H) \
+ })
+
+
+#define SHUFFLE_PATTERN_16_A__ 0x00
+#define SHUFFLE_PATTERN_16_B__ 0x01
+#define SHUFFLE_PATTERN_16_C__ 0x02
+#define SHUFFLE_PATTERN_16_D__ 0x03
+#define SHUFFLE_PATTERN_16_E__ 0x04
+#define SHUFFLE_PATTERN_16_F__ 0x05
+#define SHUFFLE_PATTERN_16_G__ 0x06
+#define SHUFFLE_PATTERN_16_H__ 0x07
+#define SHUFFLE_PATTERN_16_I__ 0x08
+#define SHUFFLE_PATTERN_16_J__ 0x09
+#define SHUFFLE_PATTERN_16_K__ 0x0a
+#define SHUFFLE_PATTERN_16_L__ 0x0b
+#define SHUFFLE_PATTERN_16_M__ 0x0c
+#define SHUFFLE_PATTERN_16_N__ 0x0d
+#define SHUFFLE_PATTERN_16_O__ 0x0e
+#define SHUFFLE_PATTERN_16_P__ 0x0f
+#define SHUFFLE_PATTERN_16_a__ 0x10
+#define SHUFFLE_PATTERN_16_b__ 0x11
+#define SHUFFLE_PATTERN_16_c__ 0x12
+#define SHUFFLE_PATTERN_16_d__ 0x13
+#define SHUFFLE_PATTERN_16_e__ 0x14
+#define SHUFFLE_PATTERN_16_f__ 0x15
+#define SHUFFLE_PATTERN_16_g__ 0x16
+#define SHUFFLE_PATTERN_16_h__ 0x17
+#define SHUFFLE_PATTERN_16_i__ 0x18
+#define SHUFFLE_PATTERN_16_j__ 0x19
+#define SHUFFLE_PATTERN_16_k__ 0x1a
+#define SHUFFLE_PATTERN_16_l__ 0x1b
+#define SHUFFLE_PATTERN_16_m__ 0x1c
+#define SHUFFLE_PATTERN_16_n__ 0x1d
+#define SHUFFLE_PATTERN_16_o__ 0x1e
+#define SHUFFLE_PATTERN_16_p__ 0x1f
+#define SHUFFLE_PATTERN_16_X__ 0xc0
+#define SHUFFLE_PATTERN_16_x__ 0xc0
+#define SHUFFLE_PATTERN_16_0__ 0x80
+#define SHUFFLE_PATTERN_16_8__ 0xe0
+
+#define SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \
+ SHUFFLE_PATTERN_16_##A##__, \
+ SHUFFLE_PATTERN_16_##B##__, \
+ SHUFFLE_PATTERN_16_##C##__, \
+ SHUFFLE_PATTERN_16_##D##__, \
+ SHUFFLE_PATTERN_16_##E##__, \
+ SHUFFLE_PATTERN_16_##F##__, \
+ SHUFFLE_PATTERN_16_##G##__, \
+ SHUFFLE_PATTERN_16_##H##__, \
+ SHUFFLE_PATTERN_16_##I##__, \
+ SHUFFLE_PATTERN_16_##J##__, \
+ SHUFFLE_PATTERN_16_##K##__, \
+ SHUFFLE_PATTERN_16_##L##__, \
+ SHUFFLE_PATTERN_16_##M##__, \
+ SHUFFLE_PATTERN_16_##N##__, \
+ SHUFFLE_PATTERN_16_##O##__, \
+ SHUFFLE_PATTERN_16_##P##__
+
+#define SHUFFLE16(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \
+ ((const vector unsigned char){ \
+ SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \
+ })
+
+#define SHUFB16(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \
+ ((const qword){ \
+ SHUFFLE_VECTOR_16__(A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P) \
+ })
+
+#endif
diff --git a/src/gallium/drivers/cell/spu/spu_texture.c b/src/gallium/drivers/cell/spu/spu_texture.c
new file mode 100644
index 00000000000..69784c89788
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_texture.c
@@ -0,0 +1,641 @@
+/**************************************************************************
+ *
+ * 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.
+ *
+ **************************************************************************/
+
+
+#include <math.h>
+
+#include "pipe/p_compiler.h"
+#include "spu_main.h"
+#include "spu_texture.h"
+#include "spu_tile.h"
+#include "spu_colorpack.h"
+#include "spu_dcache.h"
+
+
+/**
+ * Mark all tex cache entries as invalid.
+ */
+void
+invalidate_tex_cache(void)
+{
+ uint lvl;
+ for (lvl = 0; lvl < CELL_MAX_TEXTURE_LEVELS; lvl++) {
+ uint unit = 0;
+ uint bytes = 4 * spu.texture[unit].level[lvl].width
+ * spu.texture[unit].level[lvl].height;
+
+ if (spu.texture[unit].target == PIPE_TEXTURE_CUBE)
+ bytes *= 6;
+ else if (spu.texture[unit].target == PIPE_TEXTURE_3D)
+ bytes *= spu.texture[unit].level[lvl].depth;
+
+ spu_dcache_mark_dirty((unsigned) spu.texture[unit].level[lvl].start, bytes);
+ }
+}
+
+
+/**
+ * Get four texels from locations (x[0], y[0]), (x[1], y[1]) ...
+ *
+ * NOTE: in the typical case of bilinear filtering, the four texels
+ * are in a 2x2 group so we could get by with just two dcache fetches
+ * (two side-by-side texels per fetch). But when bilinear filtering
+ * wraps around a texture edge, we'll probably need code like we have
+ * now.
+ * FURTHERMORE: since we're rasterizing a quad of 2x2 pixels at a time,
+ * it's quite likely that the four pixels in a quad will need some of the
+ * same texels. So look into doing texture fetches for four pixels at
+ * a time.
+ */
+static void
+get_four_texels(const struct spu_texture_level *tlevel, uint face,
+ vec_int4 x, vec_int4 y,
+ vec_uint4 *texels)
+{
+ unsigned texture_ea = (uintptr_t) tlevel->start;
+ const vec_int4 tile_x = spu_rlmask(x, -5); /* tile_x = x / 32 */
+ const vec_int4 tile_y = spu_rlmask(y, -5); /* tile_y = y / 32 */
+ const qword offset_x = si_andi((qword) x, 0x1f); /* offset_x = x & 0x1f */
+ const qword offset_y = si_andi((qword) y, 0x1f); /* offset_y = y & 0x1f */
+
+ const qword tiles_per_row = (qword) spu_splats(tlevel->tiles_per_row);
+ const qword tile_size = (qword) spu_splats((unsigned) sizeof(tile_t));
+
+ qword tile_offset = si_mpya((qword) tile_y, tiles_per_row, (qword) tile_x);
+ tile_offset = si_mpy((qword) tile_offset, tile_size);
+
+ qword texel_offset = si_a(si_mpyui(offset_y, 32), offset_x);
+ texel_offset = si_mpyui(texel_offset, 4);
+
+ vec_uint4 offset = (vec_uint4) si_a(tile_offset, texel_offset);
+
+ texture_ea = texture_ea + face * tlevel->bytes_per_image;
+
+ spu_dcache_fetch_unaligned((qword *) & texels[0],
+ texture_ea + spu_extract(offset, 0), 4);
+ spu_dcache_fetch_unaligned((qword *) & texels[1],
+ texture_ea + spu_extract(offset, 1), 4);
+ spu_dcache_fetch_unaligned((qword *) & texels[2],
+ texture_ea + spu_extract(offset, 2), 4);
+ spu_dcache_fetch_unaligned((qword *) & texels[3],
+ texture_ea + spu_extract(offset, 3), 4);
+}
+
+
+/** clamp vec to [0, max] */
+static INLINE vector signed int
+spu_clamp(vector signed int vec, vector signed int max)
+{
+ static const vector signed int zero = {0,0,0,0};
+ vector unsigned int c;
+ c = spu_cmpgt(vec, zero); /* c = vec > zero ? ~0 : 0 */
+ vec = spu_sel(zero, vec, c);
+ c = spu_cmpgt(vec, max); /* c = vec > max ? ~0 : 0 */
+ vec = spu_sel(vec, max, c);
+ return vec;
+}
+
+
+
+/**
+ * Do nearest texture sampling for four pixels.
+ * \param colors returned colors in SOA format (rrrr, gggg, bbbb, aaaa).
+ */
+void
+sample_texture_2d_nearest(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4])
+{
+ const struct spu_texture_level *tlevel = &spu.texture[unit].level[level];
+ vector float ss = spu_mul(s, tlevel->scale_s);
+ vector float tt = spu_mul(t, tlevel->scale_t);
+ vector signed int is = spu_convts(ss, 0);
+ vector signed int it = spu_convts(tt, 0);
+ vec_uint4 texels[4];
+
+ /* PIPE_TEX_WRAP_REPEAT */
+ is = spu_and(is, tlevel->mask_s);
+ it = spu_and(it, tlevel->mask_t);
+
+ /* PIPE_TEX_WRAP_CLAMP */
+ is = spu_clamp(is, tlevel->max_s);
+ it = spu_clamp(it, tlevel->max_t);
+
+ get_four_texels(tlevel, face, is, it, texels);
+
+ /* convert four packed ARGBA pixels to float RRRR,GGGG,BBBB,AAAA */
+ spu_unpack_A8R8G8B8_transpose4(texels, colors);
+}
+
+
+/**
+ * Do bilinear texture sampling for four pixels.
+ * \param colors returned colors in SOA format (rrrr, gggg, bbbb, aaaa).
+ */
+void
+sample_texture_2d_bilinear(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4])
+{
+ const struct spu_texture_level *tlevel = &spu.texture[unit].level[level];
+ static const vector float half = {-0.5f, -0.5f, -0.5f, -0.5f};
+
+ vector float ss = spu_madd(s, tlevel->scale_s, half);
+ vector float tt = spu_madd(t, tlevel->scale_t, half);
+
+ vector signed int is0 = spu_convts(ss, 0);
+ vector signed int it0 = spu_convts(tt, 0);
+
+ /* is + 1, it + 1 */
+ vector signed int is1 = spu_add(is0, 1);
+ vector signed int it1 = spu_add(it0, 1);
+
+ /* PIPE_TEX_WRAP_REPEAT */
+ is0 = spu_and(is0, tlevel->mask_s);
+ it0 = spu_and(it0, tlevel->mask_t);
+ is1 = spu_and(is1, tlevel->mask_s);
+ it1 = spu_and(it1, tlevel->mask_t);
+
+ /* PIPE_TEX_WRAP_CLAMP */
+ is0 = spu_clamp(is0, tlevel->max_s);
+ it0 = spu_clamp(it0, tlevel->max_t);
+ is1 = spu_clamp(is1, tlevel->max_s);
+ it1 = spu_clamp(it1, tlevel->max_t);
+
+ /* get packed int texels */
+ vector unsigned int texels[16];
+ get_four_texels(tlevel, face, is0, it0, texels + 0); /* upper-left */
+ get_four_texels(tlevel, face, is1, it0, texels + 4); /* upper-right */
+ get_four_texels(tlevel, face, is0, it1, texels + 8); /* lower-left */
+ get_four_texels(tlevel, face, is1, it1, texels + 12); /* lower-right */
+
+ /* convert packed int texels to float colors */
+ vector float ftexels[16];
+ spu_unpack_A8R8G8B8_transpose4(texels + 0, ftexels + 0);
+ spu_unpack_A8R8G8B8_transpose4(texels + 4, ftexels + 4);
+ spu_unpack_A8R8G8B8_transpose4(texels + 8, ftexels + 8);
+ spu_unpack_A8R8G8B8_transpose4(texels + 12, ftexels + 12);
+
+ /* Compute weighting factors in [0,1]
+ * Multiply texcoord by 1024, AND with 1023, convert back to float.
+ */
+ vector float ss1024 = spu_mul(ss, spu_splats(1024.0f));
+ vector signed int iss1024 = spu_convts(ss1024, 0);
+ iss1024 = spu_and(iss1024, 1023);
+ vector float sWeights0 = spu_convtf(iss1024, 10);
+
+ vector float tt1024 = spu_mul(tt, spu_splats(1024.0f));
+ vector signed int itt1024 = spu_convts(tt1024, 0);
+ itt1024 = spu_and(itt1024, 1023);
+ vector float tWeights0 = spu_convtf(itt1024, 10);
+
+ /* 1 - sWeight and 1 - tWeight */
+ vector float sWeights1 = spu_sub(spu_splats(1.0f), sWeights0);
+ vector float tWeights1 = spu_sub(spu_splats(1.0f), tWeights0);
+
+ /* reds, for four pixels */
+ ftexels[ 0] = spu_mul(ftexels[ 0], spu_mul(sWeights1, tWeights1)); /*ul*/
+ ftexels[ 4] = spu_mul(ftexels[ 4], spu_mul(sWeights0, tWeights1)); /*ur*/
+ ftexels[ 8] = spu_mul(ftexels[ 8], spu_mul(sWeights1, tWeights0)); /*ll*/
+ ftexels[12] = spu_mul(ftexels[12], spu_mul(sWeights0, tWeights0)); /*lr*/
+ colors[0] = spu_add(spu_add(ftexels[0], ftexels[4]),
+ spu_add(ftexels[8], ftexels[12]));
+
+ /* greens, for four pixels */
+ ftexels[ 1] = spu_mul(ftexels[ 1], spu_mul(sWeights1, tWeights1)); /*ul*/
+ ftexels[ 5] = spu_mul(ftexels[ 5], spu_mul(sWeights0, tWeights1)); /*ur*/
+ ftexels[ 9] = spu_mul(ftexels[ 9], spu_mul(sWeights1, tWeights0)); /*ll*/
+ ftexels[13] = spu_mul(ftexels[13], spu_mul(sWeights0, tWeights0)); /*lr*/
+ colors[1] = spu_add(spu_add(ftexels[1], ftexels[5]),
+ spu_add(ftexels[9], ftexels[13]));
+
+ /* blues, for four pixels */
+ ftexels[ 2] = spu_mul(ftexels[ 2], spu_mul(sWeights1, tWeights1)); /*ul*/
+ ftexels[ 6] = spu_mul(ftexels[ 6], spu_mul(sWeights0, tWeights1)); /*ur*/
+ ftexels[10] = spu_mul(ftexels[10], spu_mul(sWeights1, tWeights0)); /*ll*/
+ ftexels[14] = spu_mul(ftexels[14], spu_mul(sWeights0, tWeights0)); /*lr*/
+ colors[2] = spu_add(spu_add(ftexels[2], ftexels[6]),
+ spu_add(ftexels[10], ftexels[14]));
+
+ /* alphas, for four pixels */
+ ftexels[ 3] = spu_mul(ftexels[ 3], spu_mul(sWeights1, tWeights1)); /*ul*/
+ ftexels[ 7] = spu_mul(ftexels[ 7], spu_mul(sWeights0, tWeights1)); /*ur*/
+ ftexels[11] = spu_mul(ftexels[11], spu_mul(sWeights1, tWeights0)); /*ll*/
+ ftexels[15] = spu_mul(ftexels[15], spu_mul(sWeights0, tWeights0)); /*lr*/
+ colors[3] = spu_add(spu_add(ftexels[3], ftexels[7]),
+ spu_add(ftexels[11], ftexels[15]));
+}
+
+
+
+/**
+ * Adapted from /opt/cell/sdk/usr/spu/include/transpose_matrix4x4.h
+ */
+static INLINE void
+transpose(vector unsigned int *mOut0,
+ vector unsigned int *mOut1,
+ vector unsigned int *mOut2,
+ vector unsigned int *mOut3,
+ vector unsigned int *mIn)
+{
+ vector unsigned int abcd, efgh, ijkl, mnop; /* input vectors */
+ vector unsigned int aeim, bfjn, cgko, dhlp; /* output vectors */
+ vector unsigned int aibj, ckdl, emfn, gohp; /* intermediate vectors */
+
+ vector unsigned char shufflehi = ((vector unsigned char) {
+ 0x00, 0x01, 0x02, 0x03,
+ 0x10, 0x11, 0x12, 0x13,
+ 0x04, 0x05, 0x06, 0x07,
+ 0x14, 0x15, 0x16, 0x17});
+ vector unsigned char shufflelo = ((vector unsigned char) {
+ 0x08, 0x09, 0x0A, 0x0B,
+ 0x18, 0x19, 0x1A, 0x1B,
+ 0x0C, 0x0D, 0x0E, 0x0F,
+ 0x1C, 0x1D, 0x1E, 0x1F});
+ abcd = *(mIn+0);
+ efgh = *(mIn+1);
+ ijkl = *(mIn+2);
+ mnop = *(mIn+3);
+
+ aibj = spu_shuffle(abcd, ijkl, shufflehi);
+ ckdl = spu_shuffle(abcd, ijkl, shufflelo);
+ emfn = spu_shuffle(efgh, mnop, shufflehi);
+ gohp = spu_shuffle(efgh, mnop, shufflelo);
+
+ aeim = spu_shuffle(aibj, emfn, shufflehi);
+ bfjn = spu_shuffle(aibj, emfn, shufflelo);
+ cgko = spu_shuffle(ckdl, gohp, shufflehi);
+ dhlp = spu_shuffle(ckdl, gohp, shufflelo);
+
+ *mOut0 = aeim;
+ *mOut1 = bfjn;
+ *mOut2 = cgko;
+ *mOut3 = dhlp;
+}
+
+
+/**
+ * Bilinear filtering, using int instead of float arithmetic for computing
+ * sample weights.
+ */
+void
+sample_texture_2d_bilinear_int(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4])
+{
+ const struct spu_texture_level *tlevel = &spu.texture[unit].level[level];
+ static const vector float half = {-0.5f, -0.5f, -0.5f, -0.5f};
+
+ /* Scale texcoords by size of texture, and add half pixel bias */
+ vector float ss = spu_madd(s, tlevel->scale_s, half);
+ vector float tt = spu_madd(t, tlevel->scale_t, half);
+
+ /* convert float coords to fixed-pt coords with 7 fraction bits */
+ vector signed int is = spu_convts(ss, 7); /* XXX really need floor() here */
+ vector signed int it = spu_convts(tt, 7); /* XXX really need floor() here */
+
+ /* compute integer texel weights in [0, 127] */
+ vector signed int sWeights0 = spu_and(is, 127);
+ vector signed int tWeights0 = spu_and(it, 127);
+ vector signed int sWeights1 = spu_sub(127, sWeights0);
+ vector signed int tWeights1 = spu_sub(127, tWeights0);
+
+ /* texel coords: is0 = is / 128, it0 = is / 128 */
+ vector signed int is0 = spu_rlmask(is, -7);
+ vector signed int it0 = spu_rlmask(it, -7);
+
+ /* texel coords: i1 = is0 + 1, it1 = it0 + 1 */
+ vector signed int is1 = spu_add(is0, 1);
+ vector signed int it1 = spu_add(it0, 1);
+
+ /* PIPE_TEX_WRAP_REPEAT */
+ is0 = spu_and(is0, tlevel->mask_s);
+ it0 = spu_and(it0, tlevel->mask_t);
+ is1 = spu_and(is1, tlevel->mask_s);
+ it1 = spu_and(it1, tlevel->mask_t);
+
+ /* PIPE_TEX_WRAP_CLAMP */
+ is0 = spu_clamp(is0, tlevel->max_s);
+ it0 = spu_clamp(it0, tlevel->max_t);
+ is1 = spu_clamp(is1, tlevel->max_s);
+ it1 = spu_clamp(it1, tlevel->max_t);
+
+ /* get packed int texels */
+ vector unsigned int texels[16];
+ get_four_texels(tlevel, face, is0, it0, texels + 0); /* upper-left */
+ get_four_texels(tlevel, face, is1, it0, texels + 4); /* upper-right */
+ get_four_texels(tlevel, face, is0, it1, texels + 8); /* lower-left */
+ get_four_texels(tlevel, face, is1, it1, texels + 12); /* lower-right */
+
+ /* twiddle packed 32-bit BGRA pixels into RGBA as four unsigned ints */
+ {
+ static const unsigned char ZERO = 0x80;
+ int i;
+ for (i = 0; i < 16; i++) {
+ texels[i] = spu_shuffle(texels[i], texels[i],
+ ((vector unsigned char) {
+ ZERO, ZERO, ZERO, 1,
+ ZERO, ZERO, ZERO, 2,
+ ZERO, ZERO, ZERO, 3,
+ ZERO, ZERO, ZERO, 0}));
+ }
+ }
+
+ /* convert RGBA,RGBA,RGBA,RGBA to RRRR,GGGG,BBBB,AAAA */
+ vector unsigned int texel0, texel1, texel2, texel3, texel4, texel5, texel6, texel7,
+ texel8, texel9, texel10, texel11, texel12, texel13, texel14, texel15;
+ transpose(&texel0, &texel1, &texel2, &texel3, texels + 0);
+ transpose(&texel4, &texel5, &texel6, &texel7, texels + 4);
+ transpose(&texel8, &texel9, &texel10, &texel11, texels + 8);
+ transpose(&texel12, &texel13, &texel14, &texel15, texels + 12);
+
+ /* computed weighted colors */
+ vector unsigned int c0, c1, c2, c3, cSum;
+
+ /* red */
+ c0 = (vector unsigned int) si_mpy((qword) texel0, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/
+ c1 = (vector unsigned int) si_mpy((qword) texel4, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/
+ c2 = (vector unsigned int) si_mpy((qword) texel8, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/
+ c3 = (vector unsigned int) si_mpy((qword) texel12, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/
+ cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3));
+ colors[0] = spu_convtf(cSum, 22);
+
+ /* green */
+ c0 = (vector unsigned int) si_mpy((qword) texel1, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/
+ c1 = (vector unsigned int) si_mpy((qword) texel5, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/
+ c2 = (vector unsigned int) si_mpy((qword) texel9, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/
+ c3 = (vector unsigned int) si_mpy((qword) texel13, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/
+ cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3));
+ colors[1] = spu_convtf(cSum, 22);
+
+ /* blue */
+ c0 = (vector unsigned int) si_mpy((qword) texel2, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/
+ c1 = (vector unsigned int) si_mpy((qword) texel6, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/
+ c2 = (vector unsigned int) si_mpy((qword) texel10, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/
+ c3 = (vector unsigned int) si_mpy((qword) texel14, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/
+ cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3));
+ colors[2] = spu_convtf(cSum, 22);
+
+ /* alpha */
+ c0 = (vector unsigned int) si_mpy((qword) texel3, si_mpy((qword) sWeights1, (qword) tWeights1)); /*ul*/
+ c1 = (vector unsigned int) si_mpy((qword) texel7, si_mpy((qword) sWeights0, (qword) tWeights1)); /*ur*/
+ c2 = (vector unsigned int) si_mpy((qword) texel11, si_mpy((qword) sWeights1, (qword) tWeights0)); /*ll*/
+ c3 = (vector unsigned int) si_mpy((qword) texel15, si_mpy((qword) sWeights0, (qword) tWeights0)); /*lr*/
+ cSum = spu_add(spu_add(c0, c1), spu_add(c2, c3));
+ colors[3] = spu_convtf(cSum, 22);
+}
+
+
+
+/**
+ * Compute level of detail factor from texcoords.
+ */
+static INLINE float
+compute_lambda_2d(uint unit, vector float s, vector float t)
+{
+ uint baseLevel = 0;
+ float width = spu.texture[unit].level[baseLevel].width;
+ float height = spu.texture[unit].level[baseLevel].width;
+ float dsdx = width * (spu_extract(s, 1) - spu_extract(s, 0));
+ float dsdy = width * (spu_extract(s, 2) - spu_extract(s, 0));
+ float dtdx = height * (spu_extract(t, 1) - spu_extract(t, 0));
+ float dtdy = height * (spu_extract(t, 2) - spu_extract(t, 0));
+#if 0
+ /* ideal value */
+ float x = dsdx * dsdx + dtdx * dtdx;
+ float y = dsdy * dsdy + dtdy * dtdy;
+ float rho = x > y ? x : y;
+ rho = sqrtf(rho);
+#else
+ /* approximation */
+ dsdx = fabsf(dsdx);
+ dsdy = fabsf(dsdy);
+ dtdx = fabsf(dtdx);
+ dtdy = fabsf(dtdy);
+ float rho = (dsdx + dsdy + dtdx + dtdy) * 0.5;
+#endif
+ float lambda = logf(rho) * 1.442695f; /* compute logbase2(rho) */
+ return lambda;
+}
+
+
+/**
+ * Blend two sets of colors according to weight.
+ */
+static void
+blend_colors(vector float c0[4], const vector float c1[4], float weight)
+{
+ vector float t = spu_splats(weight);
+ vector float dc0 = spu_sub(c1[0], c0[0]);
+ vector float dc1 = spu_sub(c1[1], c0[1]);
+ vector float dc2 = spu_sub(c1[2], c0[2]);
+ vector float dc3 = spu_sub(c1[3], c0[3]);
+ c0[0] = spu_madd(dc0, t, c0[0]);
+ c0[1] = spu_madd(dc1, t, c0[1]);
+ c0[2] = spu_madd(dc2, t, c0[2]);
+ c0[3] = spu_madd(dc3, t, c0[3]);
+}
+
+
+/**
+ * Texture sampling with level of detail selection and possibly mipmap
+ * interpolation.
+ */
+void
+sample_texture_2d_lod(vector float s, vector float t,
+ uint unit, uint level_ignored, uint face,
+ vector float colors[4])
+{
+ /*
+ * Note that we're computing a lambda/lod here that's used for all
+ * four pixels in the quad.
+ */
+ float lambda = compute_lambda_2d(unit, s, t);
+
+ (void) face;
+ (void) level_ignored;
+
+ /* apply lod bias */
+ lambda += spu.sampler[unit].lod_bias;
+
+ /* clamp */
+ if (lambda < spu.sampler[unit].min_lod)
+ lambda = spu.sampler[unit].min_lod;
+ else if (lambda > spu.sampler[unit].max_lod)
+ lambda = spu.sampler[unit].max_lod;
+
+ if (lambda <= 0.0f) {
+ /* magnify */
+ spu.mag_sample_texture_2d[unit](s, t, unit, 0, face, colors);
+ }
+ else {
+ /* minify */
+ if (spu.sampler[unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) {
+ /* sample two mipmap levels and interpolate */
+ int level = (int) lambda;
+ if (level > (int) spu.texture[unit].max_level)
+ level = spu.texture[unit].max_level;
+ spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors);
+ if (spu.sampler[unit].min_img_filter == PIPE_TEX_FILTER_LINEAR) {
+ /* sample second mipmap level */
+ float weight = lambda - (float) level;
+ level++;
+ if (level <= (int) spu.texture[unit].max_level) {
+ vector float colors2[4];
+ spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors2);
+ blend_colors(colors, colors2, weight);
+ }
+ }
+ }
+ else {
+ /* sample one mipmap level */
+ int level = (int) (lambda + 0.5f);
+ if (level > (int) spu.texture[unit].max_level)
+ level = spu.texture[unit].max_level;
+ spu.min_sample_texture_2d[unit](s, t, unit, level, face, colors);
+ }
+ }
+}
+
+
+/** XXX need a SIMD version of this */
+static unsigned
+choose_cube_face(float rx, float ry, float rz, float *newS, float *newT)
+{
+ /*
+ major axis
+ direction target sc tc ma
+ ---------- ------------------------------- --- --- ---
+ +rx TEXTURE_CUBE_MAP_POSITIVE_X_EXT -rz -ry rx
+ -rx TEXTURE_CUBE_MAP_NEGATIVE_X_EXT +rz -ry rx
+ +ry TEXTURE_CUBE_MAP_POSITIVE_Y_EXT +rx +rz ry
+ -ry TEXTURE_CUBE_MAP_NEGATIVE_Y_EXT +rx -rz ry
+ +rz TEXTURE_CUBE_MAP_POSITIVE_Z_EXT +rx -ry rz
+ -rz TEXTURE_CUBE_MAP_NEGATIVE_Z_EXT -rx -ry rz
+ */
+ const float arx = fabsf(rx);
+ const float ary = fabsf(ry);
+ const float arz = fabsf(rz);
+ unsigned face;
+ float sc, tc, ma;
+
+ if (arx > ary && arx > arz) {
+ if (rx >= 0.0F) {
+ face = PIPE_TEX_FACE_POS_X;
+ sc = -rz;
+ tc = -ry;
+ ma = arx;
+ }
+ else {
+ face = PIPE_TEX_FACE_NEG_X;
+ sc = rz;
+ tc = -ry;
+ ma = arx;
+ }
+ }
+ else if (ary > arx && ary > arz) {
+ if (ry >= 0.0F) {
+ face = PIPE_TEX_FACE_POS_Y;
+ sc = rx;
+ tc = rz;
+ ma = ary;
+ }
+ else {
+ face = PIPE_TEX_FACE_NEG_Y;
+ sc = rx;
+ tc = -rz;
+ ma = ary;
+ }
+ }
+ else {
+ if (rz > 0.0F) {
+ face = PIPE_TEX_FACE_POS_Z;
+ sc = rx;
+ tc = -ry;
+ ma = arz;
+ }
+ else {
+ face = PIPE_TEX_FACE_NEG_Z;
+ sc = -rx;
+ tc = -ry;
+ ma = arz;
+ }
+ }
+
+ *newS = (sc / ma + 1.0F) * 0.5F;
+ *newT = (tc / ma + 1.0F) * 0.5F;
+
+ return face;
+}
+
+
+
+void
+sample_texture_cube(vector float s, vector float t, vector float r,
+ uint unit, vector float colors[4])
+{
+ uint p, faces[4], level = 0;
+ float newS[4], newT[4];
+
+ /* Compute cube faces referenced by the four sets of texcoords.
+ * XXX we should SIMD-ize this.
+ */
+ for (p = 0; p < 4; p++) {
+ float rx = spu_extract(s, p);
+ float ry = spu_extract(t, p);
+ float rz = spu_extract(r, p);
+ faces[p] = choose_cube_face(rx, ry, rz, &newS[p], &newT[p]);
+ }
+
+ if (faces[0] == faces[1] &&
+ faces[0] == faces[2] &&
+ faces[0] == faces[3]) {
+ /* GOOD! All four texcoords refer to the same cube face */
+ s = (vector float) {newS[0], newS[1], newS[2], newS[3]};
+ t = (vector float) {newT[0], newT[1], newT[2], newT[3]};
+ spu.sample_texture_2d[unit](s, t, unit, level, faces[0], colors);
+ }
+ else {
+ /* BAD! The four texcoords refer to different faces */
+ for (p = 0; p < 4; p++) {
+ vector float c[4];
+
+ spu.sample_texture_2d[unit](spu_splats(newS[p]), spu_splats(newT[p]),
+ unit, level, faces[p], c);
+
+ float red = spu_extract(c[0], p);
+ float green = spu_extract(c[1], p);
+ float blue = spu_extract(c[2], p);
+ float alpha = spu_extract(c[3], p);
+
+ colors[0] = spu_insert(red, colors[0], p);
+ colors[1] = spu_insert(green, colors[1], p);
+ colors[2] = spu_insert(blue, colors[2], p);
+ colors[3] = spu_insert(alpha, colors[3], p);
+ }
+ }
+}
diff --git a/src/gallium/drivers/cell/spu/spu_texture.h b/src/gallium/drivers/cell/spu/spu_texture.h
new file mode 100644
index 00000000000..7b75b007b5a
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_texture.h
@@ -0,0 +1,67 @@
+/**************************************************************************
+ *
+ * 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 SPU_TEXTURE_H
+#define SPU_TEXTURE_H
+
+
+#include "pipe/p_compiler.h"
+
+
+extern void
+invalidate_tex_cache(void);
+
+
+extern void
+sample_texture_2d_nearest(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4]);
+
+
+extern void
+sample_texture_2d_bilinear(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4]);
+
+extern void
+sample_texture_2d_bilinear_int(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4]);
+
+
+extern void
+sample_texture_2d_lod(vector float s, vector float t,
+ uint unit, uint level, uint face,
+ vector float colors[4]);
+
+
+extern void
+sample_texture_cube(vector float s, vector float t, vector float r,
+ uint unit, vector float colors[4]);
+
+
+#endif /* SPU_TEXTURE_H */
diff --git a/src/gallium/drivers/cell/spu/spu_tile.c b/src/gallium/drivers/cell/spu/spu_tile.c
new file mode 100644
index 00000000000..6905015a483
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_tile.c
@@ -0,0 +1,126 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+
+
+#include "spu_tile.h"
+#include "spu_main.h"
+
+
+/**
+ * Get tile of color or Z values from main memory, put into SPU memory.
+ */
+void
+get_tile(uint tx, uint ty, tile_t *tile, int tag, int zBuf)
+{
+ const uint offset = ty * spu.fb.width_tiles + tx;
+ const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? spu.fb.zsize : 4);
+ const ubyte *src = zBuf ? spu.fb.depth_start : spu.fb.color_start;
+
+ src += offset * bytesPerTile;
+
+ ASSERT(tx < spu.fb.width_tiles);
+ ASSERT(ty < spu.fb.height_tiles);
+ ASSERT_ALIGN16(tile);
+ /*
+ printf("get_tile: dest: %p src: 0x%x size: %d\n",
+ tile, (unsigned int) src, bytesPerTile);
+ */
+ mfc_get(tile->ui, /* dest in local memory */
+ (unsigned int) src, /* src in main memory */
+ bytesPerTile,
+ tag,
+ 0, /* tid */
+ 0 /* rid */);
+}
+
+
+/**
+ * Move tile of color or Z values from SPU memory to main memory.
+ */
+void
+put_tile(uint tx, uint ty, const tile_t *tile, int tag, int zBuf)
+{
+ const uint offset = ty * spu.fb.width_tiles + tx;
+ const uint bytesPerTile = TILE_SIZE * TILE_SIZE * (zBuf ? spu.fb.zsize : 4);
+ ubyte *dst = zBuf ? spu.fb.depth_start : spu.fb.color_start;
+
+ dst += offset * bytesPerTile;
+
+ ASSERT(tx < spu.fb.width_tiles);
+ ASSERT(ty < spu.fb.height_tiles);
+ ASSERT_ALIGN16(tile);
+ /*
+ printf("SPU %u: put_tile: src: %p dst: 0x%x size: %d\n",
+ spu.init.id,
+ tile, (unsigned int) dst, bytesPerTile);
+ */
+ mfc_put((void *) tile->ui, /* src in local memory */
+ (unsigned int) dst, /* dst in main memory */
+ bytesPerTile,
+ tag,
+ 0, /* tid */
+ 0 /* rid */);
+}
+
+
+/**
+ * For tiles whose status is TILE_STATUS_CLEAR, write solid-filled
+ * tiles back to the main framebuffer.
+ */
+void
+really_clear_tiles(uint surfaceIndex)
+{
+ const uint num_tiles = spu.fb.width_tiles * spu.fb.height_tiles;
+ uint i;
+
+ if (surfaceIndex == 0) {
+ clear_c_tile(&spu.ctile);
+
+ for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) {
+ uint tx = i % spu.fb.width_tiles;
+ uint ty = i / spu.fb.width_tiles;
+ if (spu.ctile_status[ty][tx] == TILE_STATUS_CLEAR) {
+ put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 0);
+ }
+ }
+ }
+ else {
+ clear_z_tile(&spu.ztile);
+
+ for (i = spu.init.id; i < num_tiles; i += spu.init.num_spus) {
+ uint tx = i % spu.fb.width_tiles;
+ uint ty = i / spu.fb.width_tiles;
+ if (spu.ztile_status[ty][tx] == TILE_STATUS_CLEAR)
+ put_tile(tx, ty, &spu.ctile, TAG_SURFACE_CLEAR, 1);
+ }
+ }
+
+#if 0
+ wait_on_mask(1 << TAG_SURFACE_CLEAR);
+#endif
+}
diff --git a/src/gallium/drivers/cell/spu/spu_tile.h b/src/gallium/drivers/cell/spu/spu_tile.h
new file mode 100644
index 00000000000..7bfb52be8f3
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_tile.h
@@ -0,0 +1,75 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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 SPU_TILE_H
+#define SPU_TILE_H
+
+
+#include <libmisc.h>
+#include <spu_mfcio.h>
+#include "spu_main.h"
+#include "cell/common.h"
+
+
+
+extern void
+get_tile(uint tx, uint ty, tile_t *tile, int tag, int zBuf);
+
+extern void
+put_tile(uint tx, uint ty, const tile_t *tile, int tag, int zBuf);
+
+extern void
+really_clear_tiles(uint surfaceIndex);
+
+
+static INLINE void
+clear_c_tile(tile_t *ctile)
+{
+ memset32((uint*) ctile->ui,
+ spu.fb.color_clear_value,
+ TILE_SIZE * TILE_SIZE);
+}
+
+
+static INLINE void
+clear_z_tile(tile_t *ztile)
+{
+ if (spu.fb.zsize == 2) {
+ memset16((ushort*) ztile->us,
+ spu.fb.depth_clear_value,
+ TILE_SIZE * TILE_SIZE);
+ }
+ else {
+ ASSERT(spu.fb.zsize != 0);
+ memset32((uint*) ztile->ui,
+ spu.fb.depth_clear_value,
+ TILE_SIZE * TILE_SIZE);
+ }
+}
+
+
+#endif /* SPU_TILE_H */
diff --git a/src/gallium/drivers/cell/spu/spu_tri.c b/src/gallium/drivers/cell/spu/spu_tri.c
new file mode 100644
index 00000000000..0d9fcb99970
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_tri.c
@@ -0,0 +1,809 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+/**
+ * Triangle rendering within a tile.
+ */
+
+#include <transpose_matrix4x4.h>
+#include "pipe/p_compiler.h"
+#include "pipe/p_format.h"
+#include "util/u_math.h"
+#include "spu_colorpack.h"
+#include "spu_main.h"
+#include "spu_shuffle.h"
+#include "spu_texture.h"
+#include "spu_tile.h"
+#include "spu_tri.h"
+
+
+/** Masks are uint[4] vectors with each element being 0 or 0xffffffff */
+typedef vector unsigned int mask_t;
+
+
+
+/**
+ * Simplified types taken from other parts of Gallium
+ */
+struct vertex_header {
+ vector float data[1];
+};
+
+
+
+/* XXX fix this */
+#undef CEILF
+#define CEILF(X) ((float) (int) ((X) + 0.99999f))
+
+
+#define QUAD_TOP_LEFT 0
+#define QUAD_TOP_RIGHT 1
+#define QUAD_BOTTOM_LEFT 2
+#define QUAD_BOTTOM_RIGHT 3
+#define MASK_TOP_LEFT (1 << QUAD_TOP_LEFT)
+#define MASK_TOP_RIGHT (1 << QUAD_TOP_RIGHT)
+#define MASK_BOTTOM_LEFT (1 << QUAD_BOTTOM_LEFT)
+#define MASK_BOTTOM_RIGHT (1 << QUAD_BOTTOM_RIGHT)
+#define MASK_ALL 0xf
+
+
+#define DEBUG_VERTS 0
+
+/**
+ * Triangle edge info
+ */
+struct edge {
+ union {
+ struct {
+ float dx; /**< X(v1) - X(v0), used only during setup */
+ float dy; /**< Y(v1) - Y(v0), used only during setup */
+ };
+ vec_float4 ds; /**< vector accessor for dx and dy */
+ };
+ float dxdy; /**< dx/dy */
+ float sx, sy; /**< first sample point coord */
+ int lines; /**< number of lines on this edge */
+};
+
+
+struct interp_coef
+{
+ vector float a0;
+ vector float dadx;
+ vector float dady;
+};
+
+
+/**
+ * Triangle setup info (derived from draw_stage).
+ * Also used for line drawing (taking some liberties).
+ */
+struct setup_stage {
+
+ /* Vertices are just an array of floats making up each attribute in
+ * turn. Currently fixed at 4 floats, but should change in time.
+ * Codegen will help cope with this.
+ */
+ union {
+ struct {
+ const struct vertex_header *vmin;
+ const struct vertex_header *vmid;
+ const struct vertex_header *vmax;
+ const struct vertex_header *vprovoke;
+ };
+ qword vertex_headers;
+ };
+
+ struct edge ebot;
+ struct edge etop;
+ struct edge emaj;
+
+ float oneOverArea; /* XXX maybe make into vector? */
+
+ uint facing;
+
+ uint tx, ty; /**< position of current tile (x, y) */
+
+ int cliprect_minx, cliprect_maxx, cliprect_miny, cliprect_maxy;
+
+ struct interp_coef coef[PIPE_MAX_SHADER_INPUTS];
+
+ struct {
+ vec_int4 quad; /**< [0] = row0, [1] = row1; {left[0],left[1],right[0],right[1]} */
+ int y;
+ unsigned y_flags;
+ unsigned mask; /**< mask of MASK_BOTTOM/TOP_LEFT/RIGHT bits */
+ } span;
+};
+
+
+static struct setup_stage setup;
+
+
+/**
+ * Evaluate attribute coefficients (plane equations) to compute
+ * attribute values for the four fragments in a quad.
+ * Eg: four colors will be computed (in AoS format).
+ */
+static INLINE void
+eval_coeff(uint slot, float x, float y, vector float w, vector float result[4])
+{
+ switch (spu.vertex_info.attrib[slot].interp_mode) {
+ case INTERP_CONSTANT:
+ result[QUAD_TOP_LEFT] =
+ result[QUAD_TOP_RIGHT] =
+ result[QUAD_BOTTOM_LEFT] =
+ result[QUAD_BOTTOM_RIGHT] = setup.coef[slot].a0;
+ break;
+ case INTERP_LINEAR:
+ {
+ vector float dadx = setup.coef[slot].dadx;
+ vector float dady = setup.coef[slot].dady;
+ vector float topLeft =
+ spu_add(setup.coef[slot].a0,
+ spu_add(spu_mul(spu_splats(x), dadx),
+ spu_mul(spu_splats(y), dady)));
+
+ result[QUAD_TOP_LEFT] = topLeft;
+ result[QUAD_TOP_RIGHT] = spu_add(topLeft, dadx);
+ result[QUAD_BOTTOM_LEFT] = spu_add(topLeft, dady);
+ result[QUAD_BOTTOM_RIGHT] = spu_add(spu_add(topLeft, dadx), dady);
+ }
+ break;
+ case INTERP_PERSPECTIVE:
+ {
+ vector float dadx = setup.coef[slot].dadx;
+ vector float dady = setup.coef[slot].dady;
+ vector float topLeft =
+ spu_add(setup.coef[slot].a0,
+ spu_add(spu_mul(spu_splats(x), dadx),
+ spu_mul(spu_splats(y), dady)));
+
+ vector float wInv = spu_re(w); /* 1.0 / w */
+
+ result[QUAD_TOP_LEFT] = spu_mul(topLeft, wInv);
+ result[QUAD_TOP_RIGHT] = spu_mul(spu_add(topLeft, dadx), wInv);
+ result[QUAD_BOTTOM_LEFT] = spu_mul(spu_add(topLeft, dady), wInv);
+ result[QUAD_BOTTOM_RIGHT] = spu_mul(spu_add(spu_add(topLeft, dadx), dady), wInv);
+ }
+ break;
+ case INTERP_POS:
+ case INTERP_NONE:
+ break;
+ default:
+ ASSERT(0);
+ }
+}
+
+
+/**
+ * As above, but return 4 vectors in SOA format.
+ * XXX this will all be re-written someday.
+ */
+static INLINE void
+eval_coeff_soa(uint slot, float x, float y, vector float w, vector float result[4])
+{
+ eval_coeff(slot, x, y, w, result);
+ _transpose_matrix4x4(result, result);
+}
+
+
+/** Evalute coefficients to get Z for four pixels in a quad */
+static INLINE vector float
+eval_z(float x, float y)
+{
+ const uint slot = 0;
+ const float dzdx = spu_extract(setup.coef[slot].dadx, 2);
+ const float dzdy = spu_extract(setup.coef[slot].dady, 2);
+ const float topLeft = spu_extract(setup.coef[slot].a0, 2) + x * dzdx + y * dzdy;
+ const vector float topLeftv = spu_splats(topLeft);
+ const vector float derivs = (vector float) { 0.0, dzdx, dzdy, dzdx + dzdy };
+ return spu_add(topLeftv, derivs);
+}
+
+
+/** Evalute coefficients to get W for four pixels in a quad */
+static INLINE vector float
+eval_w(float x, float y)
+{
+ const uint slot = 0;
+ const float dwdx = spu_extract(setup.coef[slot].dadx, 3);
+ const float dwdy = spu_extract(setup.coef[slot].dady, 3);
+ const float topLeft = spu_extract(setup.coef[slot].a0, 3) + x * dwdx + y * dwdy;
+ const vector float topLeftv = spu_splats(topLeft);
+ const vector float derivs = (vector float) { 0.0, dwdx, dwdy, dwdx + dwdy };
+ return spu_add(topLeftv, derivs);
+}
+
+
+/**
+ * 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
+ * should be skipped. But adding the test for that slows things down
+ * overall.
+ */
+static INLINE void
+emit_quad( int x, int y, mask_t mask)
+{
+ /* If any bits in mask are set... */
+ if (spu_extract(spu_orx(mask), 0)) {
+ const int ix = x - setup.cliprect_minx;
+ const int iy = y - setup.cliprect_miny;
+
+ spu.cur_ctile_status = TILE_STATUS_DIRTY;
+ spu.cur_ztile_status = TILE_STATUS_DIRTY;
+
+ {
+ /*
+ * Run fragment shader, execute per-fragment ops, update fb/tile.
+ */
+ vector float inputs[4*4], outputs[2*4];
+ vector float fragZ = eval_z((float) x, (float) y);
+ vector float fragW = eval_w((float) x, (float) y);
+ vector unsigned int kill_mask;
+
+ /* setup inputs */
+#if 0
+ eval_coeff_soa(1, (float) x, (float) y, fragW, inputs);
+#else
+ uint i;
+ for (i = 0; i < spu.vertex_info.num_attribs; i++) {
+ eval_coeff_soa(i+1, (float) x, (float) y, fragW, inputs + i * 4);
+ }
+#endif
+ ASSERT(spu.fragment_program);
+ ASSERT(spu.fragment_ops);
+
+ /* Execute the current fragment program */
+ kill_mask = spu.fragment_program(inputs, outputs, spu.constants);
+
+ mask = spu_andc(mask, kill_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[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);
+ }
+ }
+}
+
+
+/**
+ * Given an X or Y coordinate, return the block/quad coordinate that it
+ * belongs to.
+ */
+static INLINE int
+block(int x)
+{
+ return x & ~1;
+}
+
+
+/**
+ * Render a horizontal span of quads
+ */
+static void
+flush_spans(void)
+{
+ int minleft, maxright;
+
+ const int l0 = spu_extract(setup.span.quad, 0);
+ const int l1 = spu_extract(setup.span.quad, 1);
+ const int r0 = spu_extract(setup.span.quad, 2);
+ const int r1 = spu_extract(setup.span.quad, 3);
+
+ switch (setup.span.y_flags) {
+ case 0x3:
+ /* both odd and even lines written (both quad rows) */
+ minleft = MIN2(l0, l1);
+ maxright = MAX2(r0, r1);
+ break;
+
+ case 0x1:
+ /* only even line written (quad top row) */
+ minleft = l0;
+ maxright = r0;
+ break;
+
+ case 0x2:
+ /* only odd line written (quad bottom row) */
+ minleft = l1;
+ maxright = r1;
+ break;
+
+ default:
+ return;
+ }
+
+ /* OK, we're very likely to need the tile data now.
+ * clear or finish waiting if needed.
+ */
+ if (spu.cur_ctile_status == TILE_STATUS_GETTING) {
+ /* wait for mfc_get() to complete */
+ //printf("SPU: %u: waiting for ctile\n", spu.init.id);
+ wait_on_mask(1 << TAG_READ_TILE_COLOR);
+ spu.cur_ctile_status = TILE_STATUS_CLEAN;
+ }
+ else if (spu.cur_ctile_status == TILE_STATUS_CLEAR) {
+ //printf("SPU %u: clearing C tile %u, %u\n", spu.init.id, setup.tx, setup.ty);
+ clear_c_tile(&spu.ctile);
+ spu.cur_ctile_status = TILE_STATUS_DIRTY;
+ }
+ ASSERT(spu.cur_ctile_status != TILE_STATUS_DEFINED);
+
+ if (spu.read_depth_stencil) {
+ if (spu.cur_ztile_status == TILE_STATUS_GETTING) {
+ /* wait for mfc_get() to complete */
+ //printf("SPU: %u: waiting for ztile\n", spu.init.id);
+ wait_on_mask(1 << TAG_READ_TILE_Z);
+ spu.cur_ztile_status = TILE_STATUS_CLEAN;
+ }
+ else if (spu.cur_ztile_status == TILE_STATUS_CLEAR) {
+ //printf("SPU %u: clearing Z tile %u, %u\n", spu.init.id, setup.tx, setup.ty);
+ clear_z_tile(&spu.ztile);
+ spu.cur_ztile_status = TILE_STATUS_DIRTY;
+ }
+ ASSERT(spu.cur_ztile_status != TILE_STATUS_DEFINED);
+ }
+
+ /* XXX this loop could be moved into the above switch cases... */
+
+ /* Setup for mask calculation */
+ const vec_int4 quad_LlRr = setup.span.quad;
+ const vec_int4 quad_RrLl = spu_rlqwbyte(quad_LlRr, 8);
+ const vec_int4 quad_LLll = spu_shuffle(quad_LlRr, quad_LlRr, SHUFFLE4(A,A,B,B));
+ const vec_int4 quad_RRrr = spu_shuffle(quad_RrLl, quad_RrLl, SHUFFLE4(A,A,B,B));
+
+ const vec_int4 twos = spu_splats(2);
+
+ const int x = block(minleft);
+ vec_int4 xs = {x, x+1, x, x+1};
+
+ for (; spu_extract(xs, 0) <= block(maxright); xs += twos) {
+ /**
+ * Computes mask to indicate which pixels in the 2x2 quad are actually
+ * inside the triangle's bounds.
+ */
+
+ /* Calculate ({x,x+1,x,x+1} >= {l[0],l[0],l[1],l[1]}) */
+ const mask_t gt_LLll_xs = spu_cmpgt(quad_LLll, xs);
+ const mask_t gte_xs_LLll = spu_nand(gt_LLll_xs, gt_LLll_xs);
+
+ /* Calculate ({r[0],r[0],r[1],r[1]} > {x,x+1,x,x+1}) */
+ const mask_t gt_RRrr_xs = spu_cmpgt(quad_RRrr, xs);
+
+ /* Combine results to create mask */
+ const mask_t mask = spu_and(gte_xs_LLll, gt_RRrr_xs);
+
+ emit_quad(spu_extract(xs, 0), setup.span.y, mask);
+ }
+
+ setup.span.y = 0;
+ setup.span.y_flags = 0;
+ /* Zero right elements */
+ setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0));
+}
+
+
+#if DEBUG_VERTS
+static void
+print_vertex(const struct vertex_header *v)
+{
+ uint i;
+ fprintf(stderr, " Vertex: (%p)\n", v);
+ for (i = 0; i < spu.vertex_info.num_attribs; i++) {
+ fprintf(stderr, " %d: %f %f %f %f\n", i,
+ spu_extract(v->data[i], 0),
+ spu_extract(v->data[i], 1),
+ spu_extract(v->data[i], 2),
+ spu_extract(v->data[i], 3));
+ }
+}
+#endif
+
+
+/**
+ * Sort vertices from top to bottom.
+ * Compute area and determine front vs. back facing.
+ * Do coarse clip test against tile bounds
+ * \return FALSE if tri is totally outside tile, TRUE otherwise
+ */
+static boolean
+setup_sort_vertices(const struct vertex_header *v0,
+ const struct vertex_header *v1,
+ const struct vertex_header *v2)
+{
+ float area, sign;
+
+#if DEBUG_VERTS
+ if (spu.init.id==0) {
+ fprintf(stderr, "SPU %u: Triangle:\n", spu.init.id);
+ print_vertex(v0);
+ print_vertex(v1);
+ print_vertex(v2);
+ }
+#endif
+
+ /* determine bottom to top order of vertices */
+ {
+ /* A table of shuffle patterns for putting vertex_header pointers into
+ correct order. Quite magical. */
+ const vec_uchar16 sort_order_patterns[] = {
+ SHUFFLE4(A,B,C,C),
+ SHUFFLE4(C,A,B,C),
+ SHUFFLE4(A,C,B,C),
+ SHUFFLE4(B,C,A,C),
+ SHUFFLE4(B,A,C,C),
+ SHUFFLE4(C,B,A,C) };
+
+ /* The vertex_header pointers, packed for easy shuffling later */
+ const vec_uint4 vs = {(unsigned)v0, (unsigned)v1, (unsigned)v2};
+
+ /* Collate y values into two vectors for comparison.
+ Using only one shuffle constant! ;) */
+ const vec_float4 y_02_ = spu_shuffle(v0->data[0], v2->data[0], SHUFFLE4(0,B,b,C));
+ const vec_float4 y_10_ = spu_shuffle(v1->data[0], v0->data[0], SHUFFLE4(0,B,b,C));
+ const vec_float4 y_012 = spu_shuffle(y_02_, v1->data[0], SHUFFLE4(0,B,b,C));
+ const vec_float4 y_120 = spu_shuffle(y_10_, v2->data[0], SHUFFLE4(0,B,b,C));
+
+ /* Perform comparison: {y0,y1,y2} > {y1,y2,y0} */
+ const vec_uint4 compare = spu_cmpgt(y_012, y_120);
+ /* Compress the result of the comparison into 4 bits */
+ const vec_uint4 gather = spu_gather(compare);
+ /* Subtract one to attain the index into the LUT. Magical. */
+ const unsigned int index = spu_extract(gather, 0) - 1;
+
+ /* Load the appropriate pattern and construct the desired vector. */
+ setup.vertex_headers = (qword)spu_shuffle(vs, vs, sort_order_patterns[index]);
+
+ /* Using the result of the comparison, set sign.
+ Very magical. */
+ sign = ((si_to_uint(si_cntb((qword)gather)) == 2) ? 1.0f : -1.0f);
+ }
+
+ /* Check if triangle is completely outside the tile bounds */
+ if (spu_extract(setup.vmin->data[0], 1) > setup.cliprect_maxy)
+ return FALSE;
+ if (spu_extract(setup.vmax->data[0], 1) < setup.cliprect_miny)
+ return FALSE;
+ if (spu_extract(setup.vmin->data[0], 0) < setup.cliprect_minx &&
+ spu_extract(setup.vmid->data[0], 0) < setup.cliprect_minx &&
+ spu_extract(setup.vmax->data[0], 0) < setup.cliprect_minx)
+ return FALSE;
+ if (spu_extract(setup.vmin->data[0], 0) > setup.cliprect_maxx &&
+ spu_extract(setup.vmid->data[0], 0) > setup.cliprect_maxx &&
+ spu_extract(setup.vmax->data[0], 0) > setup.cliprect_maxx)
+ return FALSE;
+
+ setup.ebot.ds = spu_sub(setup.vmid->data[0], setup.vmin->data[0]);
+ setup.emaj.ds = spu_sub(setup.vmax->data[0], setup.vmin->data[0]);
+ setup.etop.ds = spu_sub(setup.vmax->data[0], setup.vmid->data[0]);
+
+ /*
+ * Compute triangle's area. Use 1/area to compute partial
+ * derivatives of attributes later.
+ */
+ area = setup.emaj.dx * setup.ebot.dy - setup.ebot.dx * setup.emaj.dy;
+
+ setup.oneOverArea = 1.0f / area;
+
+ /* 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);
+
+ return TRUE;
+}
+
+
+/**
+ * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
+ * The value value comes from vertex->data[slot].
+ * The result will be put into setup.coef[slot].a0.
+ * \param slot which attribute slot
+ */
+static INLINE void
+const_coeff4(uint slot)
+{
+ setup.coef[slot].dadx = (vector float) {0.0, 0.0, 0.0, 0.0};
+ setup.coef[slot].dady = (vector float) {0.0, 0.0, 0.0, 0.0};
+ setup.coef[slot].a0 = setup.vprovoke->data[slot];
+}
+
+
+/**
+ * As above, but interp setup all four vector components.
+ */
+static INLINE void
+tri_linear_coeff4(uint slot)
+{
+ const vector float vmin_d = setup.vmin->data[slot];
+ const vector float vmid_d = setup.vmid->data[slot];
+ const vector float vmax_d = setup.vmax->data[slot];
+ const vector float xxxx = spu_splats(spu_extract(setup.vmin->data[0], 0) - 0.5f);
+ const vector float yyyy = spu_splats(spu_extract(setup.vmin->data[0], 1) - 0.5f);
+
+ vector float botda = vmid_d - vmin_d;
+ vector float majda = vmax_d - vmin_d;
+
+ vector float a = spu_sub(spu_mul(spu_splats(setup.ebot.dy), majda),
+ spu_mul(botda, spu_splats(setup.emaj.dy)));
+ vector float b = spu_sub(spu_mul(spu_splats(setup.emaj.dx), botda),
+ spu_mul(majda, spu_splats(setup.ebot.dx)));
+
+ setup.coef[slot].dadx = spu_mul(a, spu_splats(setup.oneOverArea));
+ setup.coef[slot].dady = spu_mul(b, spu_splats(setup.oneOverArea));
+
+ vector float tempx = spu_mul(setup.coef[slot].dadx, xxxx);
+ vector float tempy = spu_mul(setup.coef[slot].dady, yyyy);
+
+ setup.coef[slot].a0 = spu_sub(vmin_d, spu_add(tempx, tempy));
+}
+
+
+/**
+ * Compute a0, dadx and dady for a perspective-corrected interpolant,
+ * for a triangle.
+ * We basically multiply the vertex value by 1/w before computing
+ * the plane coefficients (a0, dadx, dady).
+ * Later, when we compute the value at a particular fragment position we'll
+ * divide the interpolated value by the interpolated W at that fragment.
+ */
+static void
+tri_persp_coeff4(uint slot)
+{
+ const vector float xxxx = spu_splats(spu_extract(setup.vmin->data[0], 0) - 0.5f);
+ const vector float yyyy = spu_splats(spu_extract(setup.vmin->data[0], 1) - 0.5f);
+
+ const vector float vmin_w = spu_splats(spu_extract(setup.vmin->data[0], 3));
+ const vector float vmid_w = spu_splats(spu_extract(setup.vmid->data[0], 3));
+ const vector float vmax_w = spu_splats(spu_extract(setup.vmax->data[0], 3));
+
+ vector float vmin_d = setup.vmin->data[slot];
+ vector float vmid_d = setup.vmid->data[slot];
+ vector float vmax_d = setup.vmax->data[slot];
+
+ vmin_d = spu_mul(vmin_d, vmin_w);
+ vmid_d = spu_mul(vmid_d, vmid_w);
+ vmax_d = spu_mul(vmax_d, vmax_w);
+
+ vector float botda = vmid_d - vmin_d;
+ vector float majda = vmax_d - vmin_d;
+
+ vector float a = spu_sub(spu_mul(spu_splats(setup.ebot.dy), majda),
+ spu_mul(botda, spu_splats(setup.emaj.dy)));
+ vector float b = spu_sub(spu_mul(spu_splats(setup.emaj.dx), botda),
+ spu_mul(majda, spu_splats(setup.ebot.dx)));
+
+ setup.coef[slot].dadx = spu_mul(a, spu_splats(setup.oneOverArea));
+ setup.coef[slot].dady = spu_mul(b, spu_splats(setup.oneOverArea));
+
+ vector float tempx = spu_mul(setup.coef[slot].dadx, xxxx);
+ vector float tempy = spu_mul(setup.coef[slot].dady, yyyy);
+
+ setup.coef[slot].a0 = spu_sub(vmin_d, spu_add(tempx, tempy));
+}
+
+
+
+/**
+ * Compute the setup.coef[] array dadx, dady, a0 values.
+ * Must be called after setup.vmin,vmid,vmax,vprovoke are initialized.
+ */
+static void
+setup_tri_coefficients(void)
+{
+ uint i;
+
+ for (i = 0; i < spu.vertex_info.num_attribs; i++) {
+ switch (spu.vertex_info.attrib[i].interp_mode) {
+ case INTERP_NONE:
+ break;
+ case INTERP_CONSTANT:
+ const_coeff4(i);
+ break;
+ case INTERP_POS:
+ /* fall-through */
+ case INTERP_LINEAR:
+ tri_linear_coeff4(i);
+ break;
+ case INTERP_PERSPECTIVE:
+ tri_persp_coeff4(i);
+ break;
+ default:
+ ASSERT(0);
+ }
+ }
+}
+
+
+static void
+setup_tri_edges(void)
+{
+ float vmin_x = spu_extract(setup.vmin->data[0], 0) + 0.5f;
+ float vmid_x = spu_extract(setup.vmid->data[0], 0) + 0.5f;
+
+ float vmin_y = spu_extract(setup.vmin->data[0], 1) - 0.5f;
+ float vmid_y = spu_extract(setup.vmid->data[0], 1) - 0.5f;
+ float vmax_y = spu_extract(setup.vmax->data[0], 1) - 0.5f;
+
+ setup.emaj.sy = CEILF(vmin_y);
+ setup.emaj.lines = (int) CEILF(vmax_y - setup.emaj.sy);
+ setup.emaj.dxdy = setup.emaj.dx / setup.emaj.dy;
+ setup.emaj.sx = vmin_x + (setup.emaj.sy - vmin_y) * setup.emaj.dxdy;
+
+ setup.etop.sy = CEILF(vmid_y);
+ setup.etop.lines = (int) CEILF(vmax_y - setup.etop.sy);
+ setup.etop.dxdy = setup.etop.dx / setup.etop.dy;
+ setup.etop.sx = vmid_x + (setup.etop.sy - vmid_y) * setup.etop.dxdy;
+
+ setup.ebot.sy = CEILF(vmin_y);
+ setup.ebot.lines = (int) CEILF(vmid_y - setup.ebot.sy);
+ setup.ebot.dxdy = setup.ebot.dx / setup.ebot.dy;
+ setup.ebot.sx = vmin_x + (setup.ebot.sy - vmin_y) * setup.ebot.dxdy;
+}
+
+
+/**
+ * Render the upper or lower half of a triangle.
+ * Scissoring/cliprect is applied here too.
+ */
+static void
+subtriangle(struct edge *eleft, struct edge *eright, unsigned lines)
+{
+ const int minx = setup.cliprect_minx;
+ const int maxx = setup.cliprect_maxx;
+ const int miny = setup.cliprect_miny;
+ const int maxy = setup.cliprect_maxy;
+ int y, start_y, finish_y;
+ int sy = (int)eleft->sy;
+
+ ASSERT((int)eleft->sy == (int) eright->sy);
+
+ /* clip top/bottom */
+ start_y = sy;
+ finish_y = sy + lines;
+
+ if (start_y < miny)
+ start_y = miny;
+
+ if (finish_y > maxy)
+ finish_y = maxy;
+
+ start_y -= sy;
+ finish_y -= sy;
+
+ /*
+ _mesa_printf("%s %d %d\n", __FUNCTION__, start_y, finish_y);
+ */
+
+ for (y = start_y; y < finish_y; y++) {
+
+ /* avoid accumulating adds as floats don't have the precision to
+ * accurately iterate large triangle edges that way. luckily we
+ * can just multiply these days.
+ *
+ * this is all drowned out by the attribute interpolation anyway.
+ */
+ int left = (int)(eleft->sx + y * eleft->dxdy);
+ int right = (int)(eright->sx + y * eright->dxdy);
+
+ /* clip left/right */
+ if (left < minx)
+ left = minx;
+ if (right > maxx)
+ right = maxx;
+
+ if (left < right) {
+ int _y = sy + y;
+ if (block(_y) != setup.span.y) {
+ flush_spans();
+ setup.span.y = block(_y);
+ }
+
+ int offset = _y&1;
+ vec_int4 quad_LlRr = {left, left, right, right};
+ /* Store left and right in 0 or 1 row of quad based on offset */
+ setup.span.quad = spu_sel(quad_LlRr, setup.span.quad, spu_maskw(5<<offset));
+ setup.span.y_flags |= 1<<offset;
+ }
+ }
+
+
+ /* save the values so that emaj can be restarted:
+ */
+ eleft->sx += lines * eleft->dxdy;
+ eright->sx += lines * eright->dxdy;
+ eleft->sy += lines;
+ eright->sy += lines;
+}
+
+
+/**
+ * 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)
+{
+ setup.tx = tx;
+ setup.ty = ty;
+
+ /* set clipping bounds to tile bounds */
+ setup.cliprect_minx = tx * TILE_SIZE;
+ setup.cliprect_miny = ty * TILE_SIZE;
+ setup.cliprect_maxx = (tx + 1) * TILE_SIZE;
+ setup.cliprect_maxy = (ty + 1) * TILE_SIZE;
+
+ if (!setup_sort_vertices((struct vertex_header *) v0,
+ (struct vertex_header *) v1,
+ (struct vertex_header *) v2)) {
+ return FALSE; /* totally clipped */
+ }
+
+ setup_tri_coefficients();
+ setup_tri_edges();
+
+ setup.span.y = 0;
+ setup.span.y_flags = 0;
+ /* Zero right elements */
+ setup.span.quad = spu_shuffle(setup.span.quad, setup.span.quad, SHUFFLE4(A,B,0,0));
+
+ if (setup.oneOverArea < 0.0) {
+ /* emaj on left */
+ subtriangle( &setup.emaj, &setup.ebot, setup.ebot.lines );
+ subtriangle( &setup.emaj, &setup.etop, setup.etop.lines );
+ }
+ else {
+ /* emaj on right */
+ subtriangle( &setup.ebot, &setup.emaj, setup.ebot.lines );
+ subtriangle( &setup.etop, &setup.emaj, setup.etop.lines );
+ }
+
+ flush_spans();
+
+ return TRUE;
+}
diff --git a/src/gallium/drivers/cell/spu/spu_tri.h b/src/gallium/drivers/cell/spu/spu_tri.h
new file mode 100644
index 00000000000..aa694dd7c93
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_tri.h
@@ -0,0 +1,37 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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 SPU_TRI_H
+#define SPU_TRI_H
+
+
+extern boolean
+tri_draw(const float *v0, const float *v1, const float *v2, uint tx, uint ty);
+
+
+#endif /* SPU_TRI_H */
diff --git a/src/gallium/drivers/cell/spu/spu_util.c b/src/gallium/drivers/cell/spu/spu_util.c
new file mode 100644
index 00000000000..b8a0d4a265f
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_util.c
@@ -0,0 +1,167 @@
+
+#include "cell/common.h"
+#include "pipe/p_shader_tokens.h"
+#include "pipe/p_debug.h"
+#include "tgsi/tgsi_parse.h"
+//#include "tgsi_build.h"
+#include "tgsi/tgsi_util.h"
+
+unsigned
+tgsi_util_get_src_register_swizzle(
+ const struct tgsi_src_register *reg,
+ unsigned component )
+{
+ switch( component ) {
+ case 0:
+ return reg->SwizzleX;
+ case 1:
+ return reg->SwizzleY;
+ case 2:
+ return reg->SwizzleZ;
+ case 3:
+ return reg->SwizzleW;
+ default:
+ ASSERT( 0 );
+ }
+ return 0;
+}
+
+unsigned
+tgsi_util_get_src_register_extswizzle(
+ const struct tgsi_src_register_ext_swz *reg,
+ unsigned component )
+{
+ switch( component ) {
+ case 0:
+ return reg->ExtSwizzleX;
+ case 1:
+ return reg->ExtSwizzleY;
+ case 2:
+ return reg->ExtSwizzleZ;
+ case 3:
+ return reg->ExtSwizzleW;
+ default:
+ ASSERT( 0 );
+ }
+ return 0;
+}
+
+unsigned
+tgsi_util_get_full_src_register_extswizzle(
+ const struct tgsi_full_src_register *reg,
+ unsigned component )
+{
+ unsigned swizzle;
+
+ /*
+ * First, calculate the extended swizzle for a given channel. This will give
+ * us either a channel index into the simple swizzle or a constant 1 or 0.
+ */
+ swizzle = tgsi_util_get_src_register_extswizzle(
+ &reg->SrcRegisterExtSwz,
+ component );
+
+ ASSERT (TGSI_SWIZZLE_X == TGSI_EXTSWIZZLE_X);
+ ASSERT (TGSI_SWIZZLE_Y == TGSI_EXTSWIZZLE_Y);
+ ASSERT (TGSI_SWIZZLE_Z == TGSI_EXTSWIZZLE_Z);
+ ASSERT (TGSI_SWIZZLE_W == TGSI_EXTSWIZZLE_W);
+ ASSERT (TGSI_EXTSWIZZLE_ZERO > TGSI_SWIZZLE_W);
+ ASSERT (TGSI_EXTSWIZZLE_ONE > TGSI_SWIZZLE_W);
+
+ /*
+ * Second, calculate the simple swizzle for the unswizzled channel index.
+ * Leave the constants intact, they are not affected by the simple swizzle.
+ */
+ if( swizzle <= TGSI_SWIZZLE_W ) {
+ swizzle = tgsi_util_get_src_register_swizzle(
+ &reg->SrcRegister,
+ component );
+ }
+
+ return swizzle;
+}
+
+unsigned
+tgsi_util_get_src_register_extnegate(
+ const struct tgsi_src_register_ext_swz *reg,
+ unsigned component )
+{
+ switch( component ) {
+ case 0:
+ return reg->NegateX;
+ case 1:
+ return reg->NegateY;
+ case 2:
+ return reg->NegateZ;
+ case 3:
+ return reg->NegateW;
+ default:
+ ASSERT( 0 );
+ }
+ return 0;
+}
+
+void
+tgsi_util_set_src_register_extnegate(
+ struct tgsi_src_register_ext_swz *reg,
+ unsigned negate,
+ unsigned component )
+{
+ switch( component ) {
+ case 0:
+ reg->NegateX = negate;
+ break;
+ case 1:
+ reg->NegateY = negate;
+ break;
+ case 2:
+ reg->NegateZ = negate;
+ break;
+ case 3:
+ reg->NegateW = negate;
+ break;
+ default:
+ ASSERT( 0 );
+ }
+}
+
+unsigned
+tgsi_util_get_full_src_register_sign_mode(
+ const struct tgsi_full_src_register *reg,
+ unsigned component )
+{
+ unsigned sign_mode;
+
+ if( reg->SrcRegisterExtMod.Absolute ) {
+ /* Consider only the post-abs negation. */
+
+ if( reg->SrcRegisterExtMod.Negate ) {
+ sign_mode = TGSI_UTIL_SIGN_SET;
+ }
+ else {
+ sign_mode = TGSI_UTIL_SIGN_CLEAR;
+ }
+ }
+ else {
+ /* Accumulate the three negations. */
+
+ unsigned negate;
+
+ negate = reg->SrcRegister.Negate;
+ if( tgsi_util_get_src_register_extnegate( &reg->SrcRegisterExtSwz, component ) ) {
+ negate = !negate;
+ }
+ if( reg->SrcRegisterExtMod.Negate ) {
+ negate = !negate;
+ }
+
+ if( negate ) {
+ sign_mode = TGSI_UTIL_SIGN_TOGGLE;
+ }
+ else {
+ sign_mode = TGSI_UTIL_SIGN_KEEP;
+ }
+ }
+
+ return sign_mode;
+}
diff --git a/src/gallium/drivers/cell/spu/spu_vertex_fetch.c b/src/gallium/drivers/cell/spu/spu_vertex_fetch.c
new file mode 100644
index 00000000000..03375d84a57
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_vertex_fetch.c
@@ -0,0 +1,145 @@
+/**************************************************************************
+ *
+ * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * (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 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.
+ *
+ **************************************************************************/
+
+ /*
+ * Authors:
+ * Keith Whitwell <[email protected]>
+ * Ian Romanick <[email protected]>
+ */
+
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "spu_exec.h"
+#include "spu_vertex_shader.h"
+#include "spu_main.h"
+#include "spu_dcache.h"
+
+typedef void (*spu_fetch_func)(qword *out, const qword *in,
+ const qword *shuffle_data);
+
+
+static const qword fetch_shuffle_data[5] ALIGN16_ATTRIB = {
+ /* Shuffle used by CVT_64_FLOAT
+ */
+ {
+ 0x00, 0x01, 0x02, 0x03, 0x10, 0x11, 0x12, 0x13,
+ 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+ },
+
+ /* Shuffle used by CVT_8_USCALED and CVT_8_SSCALED
+ */
+ {
+ 0x00, 0x80, 0x80, 0x80, 0x01, 0x80, 0x80, 0x80,
+ 0x02, 0x80, 0x80, 0x80, 0x03, 0x80, 0x80, 0x80,
+ },
+
+ /* Shuffle used by CVT_16_USCALED and CVT_16_SSCALED
+ */
+ {
+ 0x00, 0x01, 0x80, 0x80, 0x02, 0x03, 0x80, 0x80,
+ 0x04, 0x05, 0x80, 0x80, 0x06, 0x07, 0x80, 0x80,
+ },
+
+ /* High value shuffle used by trans4x4.
+ */
+ {
+ 0x00, 0x01, 0x02, 0x03, 0x10, 0x11, 0x12, 0x13,
+ 0x04, 0x05, 0x06, 0x07, 0x14, 0x15, 0x16, 0x17
+ },
+
+ /* Low value shuffle used by trans4x4.
+ */
+ {
+ 0x08, 0x09, 0x0A, 0x0B, 0x18, 0x19, 0x1A, 0x1B,
+ 0x0C, 0x0D, 0x0E, 0x0F, 0x1C, 0x1D, 0x1E, 0x1F
+ }
+};
+
+
+/**
+ * Fetch vertex attributes for 'count' vertices.
+ */
+static void generic_vertex_fetch(struct spu_vs_context *draw,
+ struct spu_exec_machine *machine,
+ const unsigned *elts,
+ unsigned count)
+{
+ unsigned nr_attrs = draw->vertex_fetch.nr_attrs;
+ unsigned attr;
+
+ ASSERT(count <= 4);
+
+#if DRAW_DBG
+ printf("SPU: %s count = %u, nr_attrs = %u\n",
+ __FUNCTION__, count, nr_attrs);
+#endif
+
+ /* loop over vertex attributes (vertex shader inputs)
+ */
+ for (attr = 0; attr < nr_attrs; attr++) {
+ const unsigned pitch = draw->vertex_fetch.pitch[attr];
+ const uint64_t src = draw->vertex_fetch.src_ptr[attr];
+ const spu_fetch_func fetch = (spu_fetch_func)
+ (draw->vertex_fetch.code + draw->vertex_fetch.code_offset[attr]);
+ unsigned i;
+ unsigned idx;
+ const unsigned bytes_per_entry = draw->vertex_fetch.size[attr];
+ const unsigned quads_per_entry = (bytes_per_entry + 15) / 16;
+ qword in[2 * 4] ALIGN16_ATTRIB;
+
+
+ /* Fetch four attributes for four vertices.
+ */
+ idx = 0;
+ for (i = 0; i < count; i++) {
+ const uint64_t addr = src + (elts[i] * pitch);
+
+#if DRAW_DBG
+ printf("SPU: fetching = 0x%llx\n", addr);
+#endif
+
+ spu_dcache_fetch_unaligned(& in[idx], addr, bytes_per_entry);
+ idx += quads_per_entry;
+ }
+
+ /* Be nice and zero out any missing vertices.
+ */
+ (void) memset(& in[idx], 0, (8 - idx) * sizeof(qword));
+
+
+ /* Convert all 4 vertices to vectors of float.
+ */
+ (*fetch)(&machine->Inputs[attr].xyzw[0].q, in, fetch_shuffle_data);
+ }
+}
+
+
+void spu_update_vertex_fetch( struct spu_vs_context *draw )
+{
+ draw->vertex_fetch.fetch_func = generic_vertex_fetch;
+}
diff --git a/src/gallium/drivers/cell/spu/spu_vertex_shader.c b/src/gallium/drivers/cell/spu/spu_vertex_shader.c
new file mode 100644
index 00000000000..fbe5b34d397
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_vertex_shader.c
@@ -0,0 +1,244 @@
+/**************************************************************************
+ *
+ * Copyright 2007 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.
+ *
+ **************************************************************************/
+
+ /*
+ * Authors:
+ * Keith Whitwell <[email protected]>
+ * Brian Paul
+ * Ian Romanick <[email protected]>
+ */
+
+#include <spu_mfcio.h>
+
+#include "pipe/p_state.h"
+#include "pipe/p_shader_tokens.h"
+#include "util/u_math.h"
+#include "draw/draw_private.h"
+#include "draw/draw_context.h"
+#include "cell/common.h"
+#include "spu_vertex_shader.h"
+#include "spu_exec.h"
+#include "spu_main.h"
+
+
+#define MAX_VERTEX_SIZE ((2 + PIPE_MAX_SHADER_OUTPUTS) * 4 * sizeof(float))
+
+
+#define CLIP_RIGHT_BIT 0x01
+#define CLIP_LEFT_BIT 0x02
+#define CLIP_TOP_BIT 0x04
+#define CLIP_BOTTOM_BIT 0x08
+#define CLIP_FAR_BIT 0x10
+#define CLIP_NEAR_BIT 0x20
+
+
+static INLINE float
+dot4(const float *a, const float *b)
+{
+ return (a[0]*b[0] +
+ a[1]*b[1] +
+ a[2]*b[2] +
+ a[3]*b[3]);
+}
+
+static INLINE unsigned
+compute_clipmask(const float *clip, /*const*/ float plane[][4], unsigned nr)
+{
+ unsigned mask = 0;
+ unsigned i;
+
+ /* Do the hardwired planes first:
+ */
+ if (-clip[0] + clip[3] < 0) mask |= CLIP_RIGHT_BIT;
+ if ( clip[0] + clip[3] < 0) mask |= CLIP_LEFT_BIT;
+ if (-clip[1] + clip[3] < 0) mask |= CLIP_TOP_BIT;
+ if ( clip[1] + clip[3] < 0) mask |= CLIP_BOTTOM_BIT;
+ if (-clip[2] + clip[3] < 0) mask |= CLIP_FAR_BIT;
+ if ( clip[2] + clip[3] < 0) mask |= CLIP_NEAR_BIT;
+
+ /* Followed by any remaining ones:
+ */
+ for (i = 6; i < nr; i++) {
+ if (dot4(clip, plane[i]) < 0)
+ mask |= (1<<i);
+ }
+
+ return mask;
+}
+
+
+/**
+ * Transform vertices with the current vertex program/shader
+ * Up to four vertices can be shaded at a time.
+ * \param vbuffer the input vertex data
+ * \param elts indexes of four input vertices
+ * \param count number of vertices to shade [1..4]
+ * \param vOut array of pointers to four output vertices
+ */
+static void
+run_vertex_program(struct spu_vs_context *draw,
+ unsigned elts[4], unsigned count,
+ const uint64_t *vOut)
+{
+ struct spu_exec_machine *machine = &draw->machine;
+ unsigned int j;
+
+ ALIGN16_DECL(struct spu_exec_vector, inputs, PIPE_MAX_ATTRIBS);
+ ALIGN16_DECL(struct spu_exec_vector, outputs, PIPE_MAX_ATTRIBS);
+ const float *scale = draw->viewport.scale;
+ const float *trans = draw->viewport.translate;
+
+ ASSERT(count <= 4);
+
+ machine->Processor = TGSI_PROCESSOR_VERTEX;
+
+ ASSERT_ALIGN16(draw->constants);
+ machine->Consts = (float (*)[4]) draw->constants;
+
+ machine->Inputs = ALIGN16_ASSIGN(inputs);
+ machine->Outputs = ALIGN16_ASSIGN(outputs);
+
+ spu_vertex_fetch( draw, machine, elts, count );
+
+ /* run shader */
+ spu_exec_machine_run( machine );
+
+
+ /* store machine results */
+ for (j = 0; j < count; j++) {
+ unsigned slot;
+ float x, y, z, w;
+ unsigned char buffer[sizeof(struct vertex_header)
+ + MAX_VERTEX_SIZE] ALIGN16_ATTRIB;
+ struct vertex_header *const tmpOut =
+ (struct vertex_header *) buffer;
+ const unsigned vert_size = ROUNDUP16(sizeof(struct vertex_header)
+ + (sizeof(float) * 4
+ * draw->num_vs_outputs));
+
+ mfc_get(tmpOut, vOut[j], vert_size, TAG_VERTEX_BUFFER, 0, 0);
+ wait_on_mask(1 << TAG_VERTEX_BUFFER);
+
+
+ /* Handle attr[0] (position) specially:
+ *
+ * XXX: Computing the clipmask should be done in the vertex
+ * program as a set of DP4 instructions appended to the
+ * user-provided code.
+ */
+ x = tmpOut->clip[0] = machine->Outputs[0].xyzw[0].f[j];
+ y = tmpOut->clip[1] = machine->Outputs[0].xyzw[1].f[j];
+ z = tmpOut->clip[2] = machine->Outputs[0].xyzw[2].f[j];
+ w = tmpOut->clip[3] = machine->Outputs[0].xyzw[3].f[j];
+
+ tmpOut->clipmask = compute_clipmask(tmpOut->clip, draw->plane,
+ draw->nr_planes);
+ tmpOut->edgeflag = 1;
+
+ /* divide by w */
+ w = 1.0f / w;
+ x *= w;
+ y *= w;
+ z *= w;
+
+ /* Viewport mapping */
+ tmpOut->data[0][0] = x * scale[0] + trans[0];
+ tmpOut->data[0][1] = y * scale[1] + trans[1];
+ tmpOut->data[0][2] = z * scale[2] + trans[2];
+ tmpOut->data[0][3] = w;
+
+ /* Remaining attributes are packed into sequential post-transform
+ * vertex attrib slots.
+ */
+ for (slot = 1; slot < draw->num_vs_outputs; slot++) {
+ tmpOut->data[slot][0] = machine->Outputs[slot].xyzw[0].f[j];
+ tmpOut->data[slot][1] = machine->Outputs[slot].xyzw[1].f[j];
+ tmpOut->data[slot][2] = machine->Outputs[slot].xyzw[2].f[j];
+ tmpOut->data[slot][3] = machine->Outputs[slot].xyzw[3].f[j];
+ }
+
+ mfc_put(tmpOut, vOut[j], vert_size, TAG_VERTEX_BUFFER, 0, 0);
+ } /* loop over vertices */
+}
+
+
+unsigned char immediates[(sizeof(float) * 4 * TGSI_EXEC_NUM_IMMEDIATES) + 32]
+ ALIGN16_ATTRIB;
+
+
+void
+spu_bind_vertex_shader(struct spu_vs_context *draw,
+ struct cell_shader_info *vs)
+{
+ const unsigned immediate_addr = vs->immediates;
+ const unsigned immediate_size =
+ ROUNDUP16((sizeof(float) * 4 * vs->num_immediates)
+ + (immediate_addr & 0x0f));
+
+
+ mfc_get(immediates, immediate_addr & ~0x0f, immediate_size,
+ TAG_VERTEX_BUFFER, 0, 0);
+
+ draw->machine.Instructions = (struct tgsi_full_instruction *)
+ vs->instructions;
+ draw->machine.NumInstructions = vs->num_instructions;
+
+ draw->machine.Declarations = (struct tgsi_full_declaration *)
+ vs->declarations;
+ draw->machine.NumDeclarations = vs->num_declarations;
+
+ draw->num_vs_outputs = vs->num_outputs;
+
+ /* specify the shader to interpret/execute */
+ spu_exec_machine_init(&draw->machine,
+ PIPE_MAX_SAMPLERS,
+ NULL /*samplers*/,
+ PIPE_SHADER_VERTEX);
+
+ wait_on_mask(1 << TAG_VERTEX_BUFFER);
+
+ (void) memcpy(& draw->machine.Imms, &immediates[immediate_addr & 0x0f],
+ sizeof(float) * 4 * vs->num_immediates);
+}
+
+
+void
+spu_execute_vertex_shader(struct spu_vs_context *draw,
+ const struct cell_command_vs *vs)
+{
+ unsigned i;
+
+ (void) memcpy(draw->plane, vs->plane, sizeof(float) * 4 * vs->nr_planes);
+ draw->nr_planes = vs->nr_planes;
+ draw->vertex_fetch.nr_attrs = vs->nr_attrs;
+
+ for (i = 0; i < vs->num_elts; i += 4) {
+ const unsigned batch_size = MIN2(vs->num_elts - i, 4);
+
+ run_vertex_program(draw, & vs->elts[i], batch_size, &vs->vOut[i]);
+ }
+}
diff --git a/src/gallium/drivers/cell/spu/spu_vertex_shader.h b/src/gallium/drivers/cell/spu/spu_vertex_shader.h
new file mode 100644
index 00000000000..4c74f5e74d5
--- /dev/null
+++ b/src/gallium/drivers/cell/spu/spu_vertex_shader.h
@@ -0,0 +1,66 @@
+#ifndef SPU_VERTEX_SHADER_H
+#define SPU_VERTEX_SHADER_H
+
+#include "cell/common.h"
+#include "pipe/p_format.h"
+#include "spu_exec.h"
+
+struct spu_vs_context;
+
+typedef void (*spu_full_fetch_func)( struct spu_vs_context *draw,
+ struct spu_exec_machine *machine,
+ const unsigned *elts,
+ unsigned count );
+
+struct spu_vs_context {
+ struct pipe_viewport_state viewport;
+
+ struct {
+ uint64_t src_ptr[PIPE_MAX_ATTRIBS];
+ unsigned pitch[PIPE_MAX_ATTRIBS];
+ unsigned size[PIPE_MAX_ATTRIBS];
+ unsigned code_offset[PIPE_MAX_ATTRIBS];
+ unsigned nr_attrs;
+ boolean dirty;
+
+ spu_full_fetch_func fetch_func;
+ void *code;
+ } vertex_fetch;
+
+ /* Clip derived state:
+ */
+ float plane[12][4];
+ unsigned nr_planes;
+
+ struct spu_exec_machine machine;
+ const float (*constants)[4];
+
+ unsigned num_vs_outputs;
+};
+
+extern void spu_update_vertex_fetch(struct spu_vs_context *draw);
+
+static INLINE void spu_vertex_fetch(struct spu_vs_context *draw,
+ struct spu_exec_machine *machine,
+ const unsigned *elts,
+ unsigned count)
+{
+ if (draw->vertex_fetch.dirty) {
+ spu_update_vertex_fetch(draw);
+ draw->vertex_fetch.dirty = 0;
+ }
+
+ (*draw->vertex_fetch.fetch_func)(draw, machine, elts, count);
+}
+
+struct cell_command_vs;
+
+extern void
+spu_bind_vertex_shader(struct spu_vs_context *draw,
+ struct cell_shader_info *vs);
+
+extern void
+spu_execute_vertex_shader(struct spu_vs_context *draw,
+ const struct cell_command_vs *vs);
+
+#endif /* SPU_VERTEX_SHADER_H */