diff options
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_curbe.c | 22 | ||||
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_vs.c | 4 | ||||
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_vs_emit.c | 139 | ||||
-rw-r--r-- | src/mesa/drivers/dri/i965/gen6_vs_state.c | 24 |
4 files changed, 96 insertions, 93 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c index 7b823eb201b..877b22fec19 100644 --- a/src/mesa/drivers/dri/i965/brw_curbe.c +++ b/src/mesa/drivers/dri/i965/brw_curbe.c @@ -242,21 +242,13 @@ static void prepare_constant_buffer(struct brw_context *brw) GLuint offset = brw->curbe.vs_start * 16; GLuint nr = brw->vs.prog_data->nr_params / 4; - if (vp->use_const_buffer) { - /* Load the subset of push constants that will get used when - * we also have a pull constant buffer. - */ - for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) { - if (brw->vs.constant_map[i] != -1) { - assert(brw->vs.constant_map[i] <= nr); - memcpy(buf + offset + brw->vs.constant_map[i] * 4, - vp->program.Base.Parameters->ParameterValues[i], - 4 * sizeof(float)); - } - } - } else { - for (i = 0; i < nr; i++) { - memcpy(buf + offset + i * 4, + /* Load the subset of push constants that will get used when + * we also have a pull constant buffer. + */ + for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) { + if (brw->vs.constant_map[i] != -1) { + assert(brw->vs.constant_map[i] <= nr); + memcpy(buf + offset + brw->vs.constant_map[i] * 4, vp->program.Base.Parameters->ParameterValues[i], 4 * sizeof(float)); } diff --git a/src/mesa/drivers/dri/i965/brw_vs.c b/src/mesa/drivers/dri/i965/brw_vs.c index 4a41c7a5176..59f270d6753 100644 --- a/src/mesa/drivers/dri/i965/brw_vs.c +++ b/src/mesa/drivers/dri/i965/brw_vs.c @@ -99,8 +99,8 @@ static void do_vs_prog( struct brw_context *brw, (void) ctx; aux_size = sizeof(c.prog_data); - if (c.vp->use_const_buffer) - aux_size += c.vp->program.Base.Parameters->NumParameters; + /* constant_map */ + aux_size += c.vp->program.Base.Parameters->NumParameters; drm_intel_bo_unreference(brw->vs.prog_bo); brw->vs.prog_bo = brw_upload_cache_with_auxdata(&brw->cache, BRW_VS_PROG, diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c index 184452666a5..e1a3f333932 100644 --- a/src/mesa/drivers/dri/i965/brw_vs_emit.c +++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c @@ -143,6 +143,8 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) GLuint i, reg = 0, mrf; int attributes_in_vue; int first_reladdr_output; + int max_constant; + int constant = 0; /* Determine whether to use a real constant buffer or use a block * of GRF registers for constants. The later is faster but only @@ -181,62 +183,81 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) } - /* Vertex program parameters from curbe: + /* Assign some (probably all) of the vertex program constants to + * the push constant buffer/CURBE. + * + * There's an obvious limit to the numer of push constants equal to + * the number of register available, and that number is smaller + * than the minimum maximum number of vertex program parameters, so + * support for pull constants is required if we overflow. + * Additionally, on gen6 the number of push constants is even + * lower. + * + * When there's relative addressing, we don't know what range of + * Mesa IR registers can be accessed. And generally, when relative + * addressing is used we also have too many constants to load them + * all as push constants. So, we'll just support relative + * addressing out of the pull constant buffers, and try to load as + * many statically-accessed constants into the push constant buffer + * as we can. */ - if (c->vp->use_const_buffer) { - int max_constant = BRW_MAX_GRF - 20 - c->vp->program.Base.NumTemporaries; - int constant = 0; - - /* We've got more constants than we can load with the push - * mechanism. This is often correlated with reladdr loads where - * we should probably be using a pull mechanism anyway to avoid - * excessive reading. However, the pull mechanism is slow in - * general. So, we try to allocate as many non-reladdr-loaded - * constants through the push buffer as we can before giving up. - */ - memset(c->constant_map, -1, c->vp->program.Base.Parameters->NumParameters); - for (i = 0; - i < c->vp->program.Base.NumInstructions && constant < max_constant; - i++) { - struct prog_instruction *inst = &c->vp->program.Base.Instructions[i]; - int arg; - - for (arg = 0; arg < 3 && constant < max_constant; arg++) { - if ((inst->SrcReg[arg].File != PROGRAM_STATE_VAR && - inst->SrcReg[arg].File != PROGRAM_CONSTANT && - inst->SrcReg[arg].File != PROGRAM_UNIFORM && - inst->SrcReg[arg].File != PROGRAM_ENV_PARAM && - inst->SrcReg[arg].File != PROGRAM_LOCAL_PARAM) || - inst->SrcReg[arg].RelAddr) - continue; - - if (c->constant_map[inst->SrcReg[arg].Index] == -1) { - c->constant_map[inst->SrcReg[arg].Index] = constant++; - } + if (intel->gen >= 6) { + /* We can only load 32 regs of push constants. */ + max_constant = 32 * 2 - c->key.nr_userclip; + } else { + max_constant = BRW_MAX_GRF - 20 - c->vp->program.Base.NumTemporaries; + } + + /* constant_map maps from ParameterValues[] index to index in the + * push constant buffer, or -1 if it's only in the pull constant + * buffer. + */ + memset(c->constant_map, -1, c->vp->program.Base.Parameters->NumParameters); + for (i = 0; + i < c->vp->program.Base.NumInstructions && constant < max_constant; + i++) { + struct prog_instruction *inst = &c->vp->program.Base.Instructions[i]; + int arg; + + for (arg = 0; arg < 3 && constant < max_constant; arg++) { + if (inst->SrcReg[arg].File != PROGRAM_STATE_VAR && + inst->SrcReg[arg].File != PROGRAM_CONSTANT && + inst->SrcReg[arg].File != PROGRAM_UNIFORM && + inst->SrcReg[arg].File != PROGRAM_ENV_PARAM && + inst->SrcReg[arg].File != PROGRAM_LOCAL_PARAM) { + continue; } - } - for (i = 0; i < constant; i++) { - c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2, - (i%2) * 4), - 0, 4, 1); + if (inst->SrcReg[arg].RelAddr) { + c->vp->use_const_buffer = GL_TRUE; + continue; + } + + if (c->constant_map[inst->SrcReg[arg].Index] == -1) { + c->constant_map[inst->SrcReg[arg].Index] = constant++; + } } - reg += (constant + 1) / 2; - c->prog_data.curb_read_length = reg - 1; - /* XXX 0 causes a bug elsewhere... */ - c->prog_data.nr_params = MAX2(constant * 4, 4); } - else { - /* use a section of the GRF for constants */ - GLuint nr_params = c->vp->program.Base.Parameters->NumParameters; - for (i = 0; i < nr_params; i++) { - c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2, (i%2) * 4), 0, 4, 1); - } - reg += (nr_params + 1) / 2; - c->prog_data.curb_read_length = reg - 1; - c->prog_data.nr_params = nr_params * 4; + /* If we ran out of push constant space, then we'll also upload all + * constants through the pull constant buffer so that they can be + * accessed no matter what. For relative addressing (the common + * case) we need them all in place anyway. + */ + if (constant == max_constant) + c->vp->use_const_buffer = GL_TRUE; + + for (i = 0; i < constant; i++) { + c->regs[PROGRAM_STATE_VAR][i] = stride(brw_vec4_grf(reg + i / 2, + (i % 2) * 4), + 0, 4, 1); } + reg += (constant + 1) / 2; + c->prog_data.curb_read_length = reg - 1; + c->prog_data.nr_params = constant; + /* XXX 0 causes a bug elsewhere... */ + if (intel->gen < 6 && c->prog_data.nr_params == 0) + c->prog_data.nr_params = 4; /* Allocate input regs: */ @@ -1302,22 +1323,18 @@ get_src_reg( struct brw_vs_compile *c, case PROGRAM_UNIFORM: case PROGRAM_ENV_PARAM: case PROGRAM_LOCAL_PARAM: - if (c->vp->use_const_buffer) { - if (!relAddr && c->constant_map[index] != -1) { - assert(c->regs[PROGRAM_STATE_VAR][c->constant_map[index]].nr != 0); - return c->regs[PROGRAM_STATE_VAR][c->constant_map[index]]; - } else if (relAddr) + if (!relAddr && c->constant_map[index] != -1) { + /* Take from the push constant buffer if possible. */ + assert(c->regs[PROGRAM_STATE_VAR][c->constant_map[index]].nr != 0); + return c->regs[PROGRAM_STATE_VAR][c->constant_map[index]]; + } else { + /* Must be in the pull constant buffer then .*/ + assert(c->vp->use_const_buffer); + if (relAddr) return get_reladdr_constant(c, inst, argIndex); else return get_constant(c, inst, argIndex); } - else if (relAddr) { - return deref(c, c->regs[PROGRAM_STATE_VAR][0], index, 16); - } - else { - assert(c->regs[PROGRAM_STATE_VAR][index].nr != 0); - return c->regs[PROGRAM_STATE_VAR][index]; - } case PROGRAM_ADDRESS: assert(index == 0); return c->regs[file][index]; diff --git a/src/mesa/drivers/dri/i965/gen6_vs_state.c b/src/mesa/drivers/dri/i965/gen6_vs_state.c index e94d0c0ddbb..4ef9e2e6072 100644 --- a/src/mesa/drivers/dri/i965/gen6_vs_state.c +++ b/src/mesa/drivers/dri/i965/gen6_vs_state.c @@ -54,7 +54,7 @@ upload_vs_state(struct brw_context *brw) OUT_BATCH(0); ADVANCE_BATCH(); } else { - int params_uploaded = 0; + int params_uploaded = 0, param_regs; float *param; if (brw->vertex_program->IsNVProgram) @@ -88,20 +88,11 @@ upload_vs_state(struct brw_context *brw) params_uploaded++; } - if (vp->use_const_buffer) { - for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) { - if (brw->vs.constant_map[i] != -1) { - memcpy(param + brw->vs.constant_map[i] * 4, - vp->program.Base.Parameters->ParameterValues[i], - 4 * sizeof(float)); - params_uploaded++; - } - } - } else { - for (i = 0; i < nr_params; i++) { - memcpy(param, vp->program.Base.Parameters->ParameterValues[i], + for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) { + if (brw->vs.constant_map[i] != -1) { + memcpy(param + brw->vs.constant_map[i] * 4, + vp->program.Base.Parameters->ParameterValues[i], 4 * sizeof(float)); - param += 4; params_uploaded++; } } @@ -117,13 +108,16 @@ upload_vs_state(struct brw_context *brw) drm_intel_gem_bo_unmap_gtt(constant_bo); + param_regs = (params_uploaded + 1) / 2; + assert(param_regs <= 32); + BEGIN_BATCH(5); OUT_BATCH(CMD_3D_CONSTANT_VS_STATE << 16 | GEN6_CONSTANT_BUFFER_0_ENABLE | (5 - 2)); OUT_RELOC(constant_bo, I915_GEM_DOMAIN_RENDER, 0, /* XXX: bad domain */ - ALIGN(params_uploaded, 2) / 2 - 1); + param_regs - 1); OUT_BATCH(0); OUT_BATCH(0); OUT_BATCH(0); |