X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fcell%2Fppu%2Fcell_state_emit.c;h=424e2628a9563483695c3d5690d7aadc5272218d;hb=ef92fe85de114cb50ca4b3070d0594aade54526c;hp=031b27f11f6da05607c686e32b97f3d5203385e1;hpb=19e4222f937c9bb95d3a899dd788afb930eecaa4;p=mesa.git diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c index 031b27f11f6..424e2628a95 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -25,8 +25,10 @@ * **************************************************************************/ -#include "pipe/p_inlines.h" +#include "util/u_inlines.h" #include "util/u_memory.h" +#include "util/u_math.h" +#include "util/u_format.h" #include "cell_context.h" #include "cell_gen_fragment.h" #include "cell_state.h" @@ -76,30 +78,86 @@ lookup_fragment_ops(struct cell_context *cell) */ if (!ops) { struct spe_function spe_code_front, spe_code_back; + unsigned int facing_dependent, total_code_size; if (0) debug_printf("**** Create New Fragment Ops\n"); - /* Prepare the buffer that will hold the generated code. */ - spe_init_func(&spe_code_front, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); - spe_init_func(&spe_code_back, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + /* Prepare the buffer that will hold the generated code. The + * "0" passed in for the size means that the SPE code will + * use a default size. + */ + spe_init_func(&spe_code_front, 0); + spe_init_func(&spe_code_back, 0); - /* generate new code. Always generate new code for both front-facing + /* Generate new code. Always generate new code for both front-facing * and back-facing fragments, even if it's the same code in both * cases. */ cell_gen_fragment_function(cell, CELL_FACING_FRONT, &spe_code_front); cell_gen_fragment_function(cell, CELL_FACING_BACK, &spe_code_back); - /* alloc new fragment ops command */ - ops = CALLOC_STRUCT(cell_command_fragment_ops); + /* Make sure the code is a multiple of 8 bytes long; this is + * required to ensure that the dual pipe instruction alignment + * is correct. It's also important for the SPU unpacking, + * which assumes 8-byte boundaries. + */ + unsigned int front_code_size = spe_code_size(&spe_code_front); + while (front_code_size % 8 != 0) { + spe_lnop(&spe_code_front); + front_code_size = spe_code_size(&spe_code_front); + } + unsigned int back_code_size = spe_code_size(&spe_code_back); + while (back_code_size % 8 != 0) { + spe_lnop(&spe_code_back); + back_code_size = spe_code_size(&spe_code_back); + } + + /* Determine whether the code we generated is facing-dependent, by + * determining whether the generated code is different for the front- + * and back-facing fragments. + */ + if (front_code_size == back_code_size && memcmp(spe_code_front.store, spe_code_back.store, front_code_size) == 0) { + /* Code is identical; only need one copy. */ + facing_dependent = 0; + total_code_size = front_code_size; + } + else { + /* Code is different for front-facing and back-facing fragments. + * Need to send both copies. + */ + facing_dependent = 1; + total_code_size = front_code_size + back_code_size; + } + /* alloc new fragment ops command. Note that this structure + * has variant length based on the total code size required. + */ + ops = CALLOC_VARIANT_LENGTH_STRUCT(cell_command_fragment_ops, total_code_size); /* populate the new cell_command_fragment_ops object */ - ops->opcode = CELL_CMD_STATE_FRAGMENT_OPS; - memcpy(ops->code_front, spe_code_front.store, spe_code_size(&spe_code_front)); - memcpy(ops->code_back, spe_code_back.store, spe_code_size(&spe_code_back)); + ops->opcode[0] = CELL_CMD_STATE_FRAGMENT_OPS; + ops->total_code_size = total_code_size; + ops->front_code_index = 0; + memcpy(ops->code, spe_code_front.store, front_code_size); + if (facing_dependent) { + /* We have separate front- and back-facing code. Append the + * back-facing code to the buffer. Be careful because the code + * size is in bytes, but the buffer is of unsigned elements. + */ + ops->back_code_index = front_code_size / sizeof(spe_code_front.store[0]); + memcpy(ops->code + ops->back_code_index, spe_code_back.store, back_code_size); + } + else { + /* Use the same code for front- and back-facing fragments */ + ops->back_code_index = ops->front_code_index; + } + + /* Set the fields for the fallback case. Note that these fields + * (and the whole fallback case) will eventually go away. + */ ops->dsa = *cell->depth_stencil; ops->blend = *cell->blend; + ops->blend_color = cell->blend_color; /* insert cell_command_fragment_ops object into keymap/cache */ util_keymap_insert(cell->fragment_ops_cache, &key, ops, NULL); @@ -122,10 +180,10 @@ static void emit_state_cmd(struct cell_context *cell, uint cmd, const void *state, uint state_size) { - uint64_t *dst = (uint64_t *) - cell_batch_alloc(cell, ROUNDUP8(sizeof(uint64_t) + state_size)); + uint32_t *dst = (uint32_t *) + cell_batch_alloc16(cell, ROUNDUP16(sizeof(opcode_t) + state_size)); *dst = cmd; - memcpy(dst + 1, state, state_size); + memcpy(dst + 4, state, state_size); } @@ -139,9 +197,10 @@ cell_emit_state(struct cell_context *cell) if (cell->dirty & CELL_NEW_FRAMEBUFFER) { struct pipe_surface *cbuf = cell->framebuffer.cbufs[0]; struct pipe_surface *zbuf = cell->framebuffer.zsbuf; + STATIC_ASSERT(sizeof(struct cell_command_framebuffer) % 16 == 0); struct cell_command_framebuffer *fb - = cell_batch_alloc(cell, sizeof(*fb)); - fb->opcode = CELL_CMD_STATE_FRAMEBUFFER; + = cell_batch_alloc16(cell, sizeof(*fb)); + fb->opcode[0] = CELL_CMD_STATE_FRAMEBUFFER; fb->color_start = cell->cbuf_map[0]; fb->color_format = cbuf->format; fb->depth_start = cell->zsbuf_map; @@ -149,23 +208,25 @@ cell_emit_state(struct cell_context *cell) fb->width = cell->framebuffer.width; fb->height = cell->framebuffer.height; #if 0 - printf("EMIT color format %s\n", pf_name(fb->color_format)); - printf("EMIT depth format %s\n", pf_name(fb->depth_format)); + printf("EMIT color format %s\n", util_format_name(fb->color_format)); + printf("EMIT depth format %s\n", util_format_name(fb->depth_format)); #endif } if (cell->dirty & (CELL_NEW_RASTERIZER)) { + STATIC_ASSERT(sizeof(struct cell_command_rasterizer) % 16 == 0); struct cell_command_rasterizer *rast = - cell_batch_alloc(cell, sizeof(*rast)); - rast->opcode = CELL_CMD_STATE_RASTERIZER; + cell_batch_alloc16(cell, sizeof(*rast)); + rast->opcode[0] = CELL_CMD_STATE_RASTERIZER; rast->rasterizer = *cell->rasterizer; } if (cell->dirty & (CELL_NEW_FS)) { /* Send new fragment program to SPUs */ + STATIC_ASSERT(sizeof(struct cell_command_fragment_program) % 16 == 0); struct cell_command_fragment_program *fp - = cell_batch_alloc(cell, sizeof(*fp)); - fp->opcode = CELL_CMD_STATE_FRAGMENT_PROGRAM; + = cell_batch_alloc16(cell, sizeof(*fp)); + fp->opcode[0] = CELL_CMD_STATE_FRAGMENT_PROGRAM; fp->num_inst = cell->fs->code.num_inst; memcpy(&fp->code, cell->fs->code.store, SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE); @@ -180,29 +241,27 @@ cell_emit_state(struct cell_context *cell) if (cell->dirty & (CELL_NEW_FS_CONSTANTS)) { const uint shader = PIPE_SHADER_FRAGMENT; - const uint num_const = cell->constants[shader].size / sizeof(float); + const uint num_const = cell->constants[shader]->size / sizeof(float); uint i, j; - float *buf = cell_batch_alloc(cell, 16 + num_const * sizeof(float)); - uint64_t *ibuf = (uint64_t *) buf; - const float *constants = pipe_buffer_map(cell->pipe.screen, - cell->constants[shader].buffer, - PIPE_BUFFER_USAGE_CPU_READ); + float *buf = cell_batch_alloc16(cell, ROUNDUP16(32 + num_const * sizeof(float))); + uint32_t *ibuf = (uint32_t *) buf; + const float *constants = cell->mapped_constants[shader]; ibuf[0] = CELL_CMD_STATE_FS_CONSTANTS; - ibuf[1] = num_const; - j = 4; + ibuf[4] = num_const; + j = 8; for (i = 0; i < num_const; i++) { buf[j++] = constants[i]; } - pipe_buffer_unmap(cell->pipe.screen, cell->constants[shader].buffer); } if (cell->dirty & (CELL_NEW_FRAMEBUFFER | CELL_NEW_DEPTH_STENCIL | CELL_NEW_BLEND)) { struct cell_command_fragment_ops *fops, *fops_cmd; - fops_cmd = cell_batch_alloc(cell, sizeof(*fops_cmd)); + /* Note that cell_command_fragment_ops is a variant-sized record */ fops = lookup_fragment_ops(cell); - memcpy(fops_cmd, fops, sizeof(*fops)); + fops_cmd = cell_batch_alloc16(cell, ROUNDUP16(sizeof(*fops_cmd) + fops->total_code_size)); + memcpy(fops_cmd, fops, sizeof(*fops) + fops->total_code_size); } if (cell->dirty & CELL_NEW_SAMPLER) { @@ -210,9 +269,10 @@ cell_emit_state(struct cell_context *cell) for (i = 0; i < CELL_MAX_SAMPLERS; i++) { if (cell->dirty_samplers & (1 << i)) { if (cell->sampler[i]) { + STATIC_ASSERT(sizeof(struct cell_command_sampler) % 16 == 0); struct cell_command_sampler *sampler - = cell_batch_alloc(cell, sizeof(*sampler)); - sampler->opcode = CELL_CMD_STATE_SAMPLER; + = cell_batch_alloc16(cell, sizeof(*sampler)); + sampler->opcode[0] = CELL_CMD_STATE_SAMPLER; sampler->unit = i; sampler->state = *cell->sampler[i]; } @@ -225,19 +285,24 @@ cell_emit_state(struct cell_context *cell) uint i; for (i = 0;i < CELL_MAX_SAMPLERS; i++) { if (cell->dirty_textures & (1 << i)) { - struct cell_command_texture *texture - = cell_batch_alloc(cell, sizeof(*texture)); - texture->opcode = CELL_CMD_STATE_TEXTURE; + STATIC_ASSERT(sizeof(struct cell_command_texture) % 16 == 0); + struct cell_command_texture *texture = + (struct cell_command_texture *) + cell_batch_alloc16(cell, sizeof(*texture)); + + texture->opcode[0] = CELL_CMD_STATE_TEXTURE; texture->unit = i; if (cell->texture[i]) { + struct cell_texture *ct = cell->texture[i]; uint level; for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) { - texture->start[level] = cell->texture[i]->tiled_mapped[level]; - texture->width[level] = cell->texture[i]->base.width[level]; - texture->height[level] = cell->texture[i]->base.height[level]; - texture->depth[level] = cell->texture[i]->base.depth[level]; + texture->start[level] = (ct->mapped + + ct->level_offset[level]); + texture->width[level] = u_minify(ct->base.width0, level); + texture->height[level] = u_minify(ct->base.height0, level); + texture->depth[level] = u_minify(ct->base.depth0, level); } - texture->target = cell->texture[i]->base.target; + texture->target = ct->base.target; } else { uint level; @@ -264,7 +329,7 @@ cell_emit_state(struct cell_context *cell) const struct draw_context *const draw = cell->draw; struct cell_shader_info info; - info.num_outputs = draw_num_vs_outputs(draw); + info.num_outputs = draw_num_shader_outputs(draw); info.declarations = (uintptr_t) draw->vs.machine.Declarations; info.num_declarations = draw->vs.machine.NumDeclarations; info.instructions = (uintptr_t) draw->vs.machine.Instructions;