X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fcell%2Fspu%2Fspu_command.c;h=5c0179d95462015ce358303016003bdfb0f84ac3;hb=4bced42341d13303ae023957d3001a640cf7ea2d;hp=4e98eea3389a15090c31bb87770940ad4b361650;hpb=f8bddf698d523f597fea0f721b064daee81d8005;p=mesa.git diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c index 4e98eea3389..5c0179d9546 100644 --- a/src/gallium/drivers/cell/spu/spu_command.c +++ b/src/gallium/drivers/cell/spu/spu_command.c @@ -44,7 +44,6 @@ #include "spu_tile.h" #include "spu_vertex_shader.h" #include "spu_dcache.h" -#include "spu_debug.h" #include "cell/common.h" @@ -59,6 +58,14 @@ static unsigned char attribute_fetch_code_buffer[136 * PIPE_MAX_ATTRIBS] +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. @@ -69,9 +76,10 @@ static void release_buffer(uint buffer) { /* Evidently, using less than a 16-byte status doesn't work reliably */ - static const uint status[4] ALIGN16_ATTRIB - = {CELL_BUFFER_STATUS_FREE, 0, 0, 0}; - + 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; @@ -86,10 +94,33 @@ release_buffer(uint buffer) } +/** + * 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) { - DEBUG_PRINTF("CLEAR SURF %u to 0x%08x\n", clear->surface, clear->value); + 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; @@ -157,14 +188,14 @@ cmd_clear_surface(const struct cell_command_clear_surface *clear) #endif /* CLEAR_OPT */ - DEBUG_PRINTF("CLEAR SURF done\n"); + D_PRINTF(CELL_DEBUG_CMD, "CLEAR SURF done\n"); } static void cmd_release_verts(const struct cell_command_release_verts *release) { - DEBUG_PRINTF("RELEASE VERTS %u\n", release->vertex_buf); + D_PRINTF(CELL_DEBUG_CMD, "RELEASE VERTS %u\n", release->vertex_buf); ASSERT(release->vertex_buf != ~0U); release_buffer(release->vertex_buf); } @@ -179,48 +210,77 @@ cmd_release_verts(const struct cell_command_release_verts *release) static void cmd_state_fragment_ops(const struct cell_command_fragment_ops *fops) { - static int warned = 0; + D_PRINTF(CELL_DEBUG_CMD, "CMD_STATE_FRAGMENT_OPS\n"); - DEBUG_PRINTF("CMD_STATE_FRAGMENT_OPS\n"); - /* Copy SPU code from batch buffer to spu buffer */ - memcpy(spu.fragment_ops_code, fops->code, SPU_MAX_FRAGMENT_OPS_INSTS * 4); - /* Copy state info (for fallback case only) */ + /* 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)); - /* Parity twist! For now, always use the fallback code by default, - * only switching to codegen when specifically requested. This - * allows us to develop freely without risking taking down the - * branch. - * - * Later, the parity of this check will be reversed, so that - * codegen is *always* used, unless we specifically indicate that - * we don't want it. - * - * Eventually, the option will be removed completely, because in - * final code we'll always use codegen and won't even provide the - * raw state records that the fallback code requires. + /* Make sure the SPU knows which buffers it's expected to read when + * it's told to pull tiles. */ - if ((spu.init.debug_flags & CELL_DEBUG_FRAGMENT_OP_FALLBACK) == 0) { - spu.fragment_ops = (spu_fragment_ops_func) spu.fragment_ops_code; - } - else { - /* otherwise, the default fallback code remains in place */ + 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; } - spu.read_depth = spu.depth_stencil_alpha.depth.enabled; - spu.read_stencil = spu.depth_stencil_alpha.stencil[0].enabled; -} + /* 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) { - DEBUG_PRINTF("CMD_STATE_FRAGMENT_PROGRAM\n"); + 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); @@ -232,28 +292,29 @@ cmd_state_fragment_program(const struct cell_command_fragment_program *fp) static uint -cmd_state_fs_constants(const uint64_t *buffer, uint pos) +cmd_state_fs_constants(const qword *buffer, uint pos) { - const uint num_const = buffer[pos + 1]; - const float *constants = (const float *) &buffer[pos + 2]; + const uint num_const = spu_extract((vector unsigned int)buffer[pos+1], 0); + const float *constants = (const float *) &buffer[pos+2]; uint i; - DEBUG_PRINTF("CMD_STATE_FS_CONSTANTS (%u)\n", num_const); + 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 8-byte words) */ - return pos + 2 + num_const / 2; + /* 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) { - DEBUG_PRINTF("FRAMEBUFFER: %d x %d at %p, cformat 0x%x zformat 0x%x\n", + 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, @@ -295,26 +356,90 @@ cmd_state_framebuffer(const struct cell_command_framebuffer *cmd) } +/** + * 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) { - DEBUG_PRINTF("SAMPLER [%u]\n", sampler->unit); + uint unit = sampler->unit; + + D_PRINTF(CELL_DEBUG_CMD, "SAMPLER [%u]\n", unit); - spu.sampler[sampler->unit] = sampler->state; + spu.sampler[unit] = sampler->state; - if (spu.sampler[sampler->unit].min_mip_filter != PIPE_TEX_MIPFILTER_NONE) { - /* use lambda/lod to determine min vs. mag filter */ - spu.sample_texture4[sampler->unit] = sample_texture4_lod; + 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); } - else if (spu.sampler[sampler->unit].min_img_filter - == PIPE_TEX_FILTER_LINEAR) { - /* min = mag = bilinear */ - spu.sample_texture4[sampler->unit] = sample_texture4_bilinear; + + 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); } - else { - /* min = mag = inearest */ - spu.sample_texture4[sampler->unit] = sample_texture4_nearest; + + 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]); } @@ -324,43 +449,45 @@ cmd_state_texture(const struct cell_command_texture *texture) const uint unit = texture->unit; uint i; - //if (spu.init.id==0) Debug=1; - - DEBUG_PRINTF("TEXTURE [%u]\n", texture->unit); + 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]; - DEBUG_PRINTF(" LEVEL %u: at %p size[0] %u x %u\n", 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].width4 = spu_splats((float) width); - spu.texture[unit].level[i].height4 = spu_splats((float) height); + spu.texture[unit].level[i].bytes_per_image = + 4 * align(width, TILE_SIZE) * align(height, TILE_SIZE) * depth; - spu.texture[unit].level[i].tex_size_x_mask = spu_splats(width - 1); - spu.texture[unit].level[i].tex_size_y_mask = spu_splats(height - 1); + 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; } - //Debug=0; + + update_tex_masks(&spu.texture[unit], &spu.sampler[unit]); } static void cmd_state_vertex_info(const struct vertex_info *vinfo) { - DEBUG_PRINTF("VERTEX_INFO num_attribs=%u\n", vinfo->num_attribs); + 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)); @@ -399,7 +526,7 @@ cmd_state_attrib_fetch(const struct cell_attribute_fetch_code *code) static void cmd_finish(void) { - DEBUG_PRINTF("FINISH\n"); + D_PRINTF(CELL_DEBUG_CMD, "FINISH\n"); really_clear_tiles(0); /* wait for all outstanding DMAs to finish */ mfc_write_tag_mask(~0); @@ -420,11 +547,11 @@ cmd_batch(uint opcode) { const uint buf = (opcode >> 8) & 0xff; uint size = (opcode >> 16); - uint64_t buffer[CELL_BUFFER_SIZE / 8] ALIGN16_ATTRIB; - const unsigned usize = size / sizeof(buffer[0]); + qword buffer[CELL_BUFFER_SIZE / 16] ALIGN16_ATTRIB; + const unsigned usize = ROUNDUP16(size) / sizeof(buffer[0]); uint pos; - DEBUG_PRINTF("BATCH buffer %u, len %u, from %p\n", + 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); @@ -444,14 +571,14 @@ cmd_batch(uint opcode) wait_on_mask(1 << TAG_BATCH_BUFFER); /* Tell PPU we're done copying the buffer to local store */ - DEBUG_PRINTF("release batch buf %u\n", buf); + 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 (buffer[pos]) { + switch (si_to_uint(buffer[pos])) { /* * rendering commands */ @@ -460,7 +587,7 @@ cmd_batch(uint opcode) struct cell_command_clear_surface *clr = (struct cell_command_clear_surface *) &buffer[pos]; cmd_clear_surface(clr); - pos += sizeof(*clr) / 8; + pos += sizeof(*clr) / 16; } break; case CELL_CMD_RENDER: @@ -469,7 +596,7 @@ cmd_batch(uint opcode) = (struct cell_command_render *) &buffer[pos]; uint pos_incr; cmd_render(render, &pos_incr); - pos += pos_incr; + pos += ((pos_incr+1)&~1) / 2; // should 'fix' cmd_render return } break; /* @@ -480,7 +607,7 @@ cmd_batch(uint opcode) struct cell_command_framebuffer *fb = (struct cell_command_framebuffer *) &buffer[pos]; cmd_state_framebuffer(fb); - pos += sizeof(*fb) / 8; + pos += sizeof(*fb) / 16; } break; case CELL_CMD_STATE_FRAGMENT_OPS: @@ -488,7 +615,8 @@ cmd_batch(uint opcode) struct cell_command_fragment_ops *fops = (struct cell_command_fragment_ops *) &buffer[pos]; cmd_state_fragment_ops(fops); - pos += sizeof(*fops) / 8; + /* This is a variant-sized command */ + pos += ROUNDUP16(sizeof(*fops) + fops->total_code_size) / 16; } break; case CELL_CMD_STATE_FRAGMENT_PROGRAM: @@ -496,18 +624,26 @@ cmd_batch(uint opcode) struct cell_command_fragment_program *fp = (struct cell_command_fragment_program *) &buffer[pos]; cmd_state_fragment_program(fp); - pos += sizeof(*fp) / 8; + 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) / 8; + pos += sizeof(*sampler) / 16; } break; case CELL_CMD_STATE_TEXTURE: @@ -515,37 +651,37 @@ cmd_batch(uint opcode) struct cell_command_texture *texture = (struct cell_command_texture *) &buffer[pos]; cmd_state_texture(texture); - pos += sizeof(*texture) / 8; + pos += sizeof(*texture) / 16; } break; case CELL_CMD_STATE_VERTEX_INFO: cmd_state_vertex_info((struct vertex_info *) &buffer[pos+1]); - pos += (1 + ROUNDUP8(sizeof(struct vertex_info)) / 8); + 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 + ROUNDUP8(sizeof(struct pipe_viewport_state)) / 8); + pos += 1 + ROUNDUP16(sizeof(struct pipe_viewport_state)) / 16; break; case CELL_CMD_STATE_UNIFORMS: - draw.constants = (const float (*)[4]) (uintptr_t) buffer[pos + 1]; + 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 + ROUNDUP8(sizeof(struct cell_array_info)) / 8); + 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 + ROUNDUP8(sizeof(struct cell_shader_info)) / 8); + 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 + ROUNDUP8(sizeof(struct cell_attribute_fetch_code)) / 8); + pos += 1 + ROUNDUP16(sizeof(struct cell_attribute_fetch_code)) / 16; break; /* * misc commands @@ -554,12 +690,20 @@ cmd_batch(uint opcode) 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) / 8; + pos += sizeof(*release) / 16; } break; case CELL_CMD_FLUSH_BUFFER_RANGE: { @@ -567,20 +711,22 @@ cmd_batch(uint opcode) &buffer[pos+1]; spu_dcache_mark_dirty((unsigned) br->base, br->size); - pos += (1 + ROUNDUP8(sizeof(struct cell_buffer_range)) / 8); + pos += 1 + ROUNDUP16(sizeof(struct cell_buffer_range)) / 16; break; } default: - printf("SPU %u: bad opcode: 0x%llx\n", spu.init.id, buffer[pos]); + printf("SPU %u: bad opcode: 0x%x\n", spu.init.id, si_to_uint(buffer[pos])); ASSERT(0); break; } } - DEBUG_PRINTF("BATCH complete\n"); + D_PRINTF(CELL_DEBUG_CMD, "BATCH complete\n"); } +#define PERF 0 + /** * Main loop for SPEs: Get a command, execute it, repeat. @@ -588,41 +734,29 @@ cmd_batch(uint opcode) void command_loop(void) { - struct cell_command cmd; int exitFlag = 0; + uint t0, t1; - DEBUG_PRINTF("Enter command loop\n"); - - ASSERT((sizeof(struct cell_command) & 0xf) == 0); - ASSERT_ALIGN16(&cmd); + D_PRINTF(CELL_DEBUG_CMD, "Enter command loop\n"); while (!exitFlag) { unsigned opcode; - int tag = 0; - DEBUG_PRINTF("Wait for cmd...\n"); + 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); - DEBUG_PRINTF("got cmd 0x%x\n", opcode); - - /* command payload */ - mfc_get(&cmd, /* dest */ - (unsigned int) spu.init.cmd, /* src */ - sizeof(struct cell_command), /* bytes */ - tag, - 0, /* tid */ - 0 /* rid */); - wait_on_mask( 1 << tag ); - - /* - * NOTE: most commands should be contained in a batch buffer - */ + if (PERF) + t0 = spu_read_decrementer(); switch (opcode & CELL_CMD_OPCODE_MASK) { case CELL_CMD_EXIT: - DEBUG_PRINTF("EXIT\n"); + D_PRINTF(CELL_DEBUG_CMD, "EXIT\n"); exitFlag = 1; break; case CELL_CMD_VS_EXECUTE: @@ -637,9 +771,45 @@ command_loop(void) 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); + } } - DEBUG_PRINTF("Exit command loop\n"); + D_PRINTF(CELL_DEBUG_CMD, "Exit command loop\n"); + + if (spu.init.debug_flags & CELL_DEBUG_CACHE) + spu_dcache_report(); +} - 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(); }