From: Samuel Iglesias Gonsálvez Date: Wed, 19 Apr 2017 09:18:35 +0000 (+0200) Subject: i965/vec4: load dvec3/4 uniforms first in the push constant buffer X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=e69e5c7006da80af62c9ef08dec215b3b4b30946;p=mesa.git i965/vec4: load dvec3/4 uniforms first in the push constant buffer Reorder the uniforms to load first the dvec4-aligned variables in the push constant buffer and then push the vec4-aligned ones. It takes into account that the relocated uniforms should be aligned to their channel size. This fixes a bug were the dvec3/4 might be loaded one part on a GRF and the rest in next GRF, so the region parameters to read that could break the HW rules. v2: - Fix broken logic. - Add a comment to explain what should be needed to optimise the usage of the push constant buffer slots, as this patch does not pack the uniforms. v3: - Implemented the push constant buffer usage optimization. Signed-off-by: Samuel Iglesias Gonsálvez Cc: "17.1" Acked-by: Francisco Jerez --- diff --git a/src/intel/compiler/brw_vec4.cpp b/src/intel/compiler/brw_vec4.cpp index 5c032c0c822..a641ebfbba9 100644 --- a/src/intel/compiler/brw_vec4.cpp +++ b/src/intel/compiler/brw_vec4.cpp @@ -583,16 +583,46 @@ vec4_visitor::split_uniform_registers() } } +/* This function returns the register number where we placed the uniform */ +static int +set_push_constant_loc(const int nr_uniforms, int *new_uniform_count, + const int src, const int size, const int channel_size, + int *new_loc, int *new_chan, + int *new_chans_used) +{ + int dst; + /* Find the lowest place we can slot this uniform in. */ + for (dst = 0; dst < nr_uniforms; dst++) { + if (ALIGN(new_chans_used[dst], channel_size) + size <= 4) + break; + } + + assert(dst < nr_uniforms); + + new_loc[src] = dst; + new_chan[src] = ALIGN(new_chans_used[dst], channel_size); + new_chans_used[dst] = ALIGN(new_chans_used[dst], channel_size) + size; + + *new_uniform_count = MAX2(*new_uniform_count, dst + 1); + return dst; +} + void vec4_visitor::pack_uniform_registers() { uint8_t chans_used[this->uniforms]; int new_loc[this->uniforms]; int new_chan[this->uniforms]; + bool is_aligned_to_dvec4[this->uniforms]; + int new_chans_used[this->uniforms]; + int channel_sizes[this->uniforms]; memset(chans_used, 0, sizeof(chans_used)); memset(new_loc, 0, sizeof(new_loc)); memset(new_chan, 0, sizeof(new_chan)); + memset(new_chans_used, 0, sizeof(new_chans_used)); + memset(is_aligned_to_dvec4, 0, sizeof(is_aligned_to_dvec4)); + memset(channel_sizes, 0, sizeof(channel_sizes)); /* Find which uniform vectors are actually used by the program. We * expect unused vector elements when we've moved array access out @@ -622,7 +652,7 @@ vec4_visitor::pack_uniform_registers() continue; assert(type_sz(inst->src[i].type) % 4 == 0); - unsigned channel_size = type_sz(inst->src[i].type) / 4; + int channel_size = type_sz(inst->src[i].type) / 4; int reg = inst->src[i].nr; for (int c = 0; c < 4; c++) { @@ -631,10 +661,15 @@ vec4_visitor::pack_uniform_registers() unsigned channel = BRW_GET_SWZ(inst->src[i].swizzle, c) + 1; unsigned used = MAX2(chans_used[reg], channel * channel_size); - if (used <= 4) + if (used <= 4) { chans_used[reg] = used; - else + channel_sizes[reg] = MAX2(channel_sizes[reg], channel_size); + } else { + is_aligned_to_dvec4[reg] = true; + is_aligned_to_dvec4[reg + 1] = true; chans_used[reg + 1] = used - 4; + channel_sizes[reg + 1] = MAX2(channel_sizes[reg + 1], channel_size); + } } } @@ -659,42 +694,60 @@ vec4_visitor::pack_uniform_registers() int new_uniform_count = 0; + /* As the uniforms are going to be reordered, take the data from a temporary + * copy of the original param[]. + */ + gl_constant_value **param = ralloc_array(NULL, gl_constant_value*, + stage_prog_data->nr_params); + memcpy(param, stage_prog_data->param, + sizeof(gl_constant_value*) * stage_prog_data->nr_params); + /* Now, figure out a packing of the live uniform vectors into our - * push constants. + * push constants. Start with dvec{3,4} because they are aligned to + * dvec4 size (2 vec4). */ for (int src = 0; src < uniforms; src++) { int size = chans_used[src]; - if (size == 0) + if (size == 0 || !is_aligned_to_dvec4[src]) continue; - int dst; - /* Find the lowest place we can slot this uniform in. */ - for (dst = 0; dst < src; dst++) { - if (chans_used[dst] + size <= 4) - break; + /* dvec3 are aligned to dvec4 size, apply the alignment of the size + * to 4 to avoid moving last component of a dvec3 to the available + * location at the end of a previous dvec3. These available locations + * could be filled by smaller variables in next loop. + */ + size = ALIGN(size, 4); + int dst = set_push_constant_loc(uniforms, &new_uniform_count, + src, size, channel_sizes[src], + new_loc, new_chan, + new_chans_used); + /* Move the references to the data */ + for (int j = 0; j < size; j++) { + stage_prog_data->param[dst * 4 + new_chan[src] + j] = + param[src * 4 + j]; } + } - if (src == dst) { - new_loc[src] = dst; - new_chan[src] = 0; - } else { - new_loc[src] = dst; - new_chan[src] = chans_used[dst]; + /* Continue with the rest of data, which is aligned to vec4. */ + for (int src = 0; src < uniforms; src++) { + int size = chans_used[src]; - /* Move the references to the data */ - for (int j = 0; j < size; j++) { - stage_prog_data->param[dst * 4 + new_chan[src] + j] = - stage_prog_data->param[src * 4 + j]; - } + if (size == 0 || is_aligned_to_dvec4[src]) + continue; - chans_used[dst] += size; - chans_used[src] = 0; + int dst = set_push_constant_loc(uniforms, &new_uniform_count, + src, size, channel_sizes[src], + new_loc, new_chan, + new_chans_used); + /* Move the references to the data */ + for (int j = 0; j < size; j++) { + stage_prog_data->param[dst * 4 + new_chan[src] + j] = + param[src * 4 + j]; } - - new_uniform_count = MAX2(new_uniform_count, dst + 1); } + ralloc_free(param); this->uniforms = new_uniform_count; /* Now, update the instructions for our repacked uniforms. */ @@ -705,9 +758,9 @@ vec4_visitor::pack_uniform_registers() if (inst->src[i].file != UNIFORM) continue; + int chan = new_chan[src] / channel_sizes[src]; inst->src[i].nr = new_loc[src]; - inst->src[i].swizzle += BRW_SWIZZLE4(new_chan[src], new_chan[src], - new_chan[src], new_chan[src]); + inst->src[i].swizzle += BRW_SWIZZLE4(chan, chan, chan, chan); } } }