From b3faba50a323244a091d85779d46448889d4bf63 Mon Sep 17 00:00:00 2001
From: Ben Vanik <ben.vanik@gmail.com>
Date: Sun, 21 Feb 2016 10:38:58 -0800
Subject: [PATCH] Pushing constants.
---
src/xenia/gpu/spirv_shader_translator.cc | 36 +++++-----
src/xenia/gpu/spirv_shader_translator.h | 22 ++++++
src/xenia/gpu/vulkan/pipeline_cache.cc | 85 ++++++++++++++++++------
src/xenia/gpu/vulkan/pipeline_cache.h | 2 +
4 files changed, 111 insertions(+), 34 deletions(-)
diff --git a/src/xenia/gpu/spirv_shader_translator.cc b/src/xenia/gpu/spirv_shader_translator.cc
index f6f2ba229..ec19f76c0 100644
--- a/src/xenia/gpu/spirv_shader_translator.cc
+++ b/src/xenia/gpu/spirv_shader_translator.cc
@@ -133,24 +133,30 @@ void SpirvShaderTranslator::StartTranslation() {
b.addDecoration(consts_, spv::Decoration::DecorationBinding, 1);
}
- // Push constants.
- Id push_constants_type =
- b.makeStructType({vec4_float_type_, vec4_float_type_, vec4_float_type_},
- "push_consts_type");
-
- b.addMemberDecoration(push_constants_type, 0,
- spv::Decoration::DecorationOffset, 0);
+ // Push constants, represented by SpirvPushConstants.
+ Id push_constants_type = b.makeStructType(
+ {vec4_float_type_, vec4_float_type_, vec4_float_type_, uint_type},
+ "push_consts_type");
+ // float4 window_scale;
+ b.addMemberDecoration(
+ push_constants_type, 0, spv::Decoration::DecorationOffset,
+ static_cast<int>(offsetof(SpirvPushConstants, window_scale)));
b.addMemberName(push_constants_type, 0, "window_scale");
-
- b.addMemberDecoration(push_constants_type, 1,
- spv::Decoration::DecorationOffset, 4 * sizeof(float));
+ // float4 vtx_fmt;
+ b.addMemberDecoration(
+ push_constants_type, 1, spv::Decoration::DecorationOffset,
+ static_cast<int>(offsetof(SpirvPushConstants, vtx_fmt)));
b.addMemberName(push_constants_type, 1, "vtx_fmt");
-
- b.addMemberDecoration(push_constants_type, 2,
- spv::Decoration::DecorationOffset,
- 2 * 4 * sizeof(float));
+ // float4 alpha_test;
+ b.addMemberDecoration(
+ push_constants_type, 2, spv::Decoration::DecorationOffset,
+ static_cast<int>(offsetof(SpirvPushConstants, alpha_test)));
b.addMemberName(push_constants_type, 2, "alpha_test");
-
+ // uint ps_param_gen;
+ b.addMemberDecoration(
+ push_constants_type, 3, spv::Decoration::DecorationOffset,
+ static_cast<int>(offsetof(SpirvPushConstants, ps_param_gen)));
+ b.addMemberName(push_constants_type, 3, "ps_param_gen");
push_consts_ = b.createVariable(spv::StorageClass::StorageClassPushConstant,
push_constants_type, "push_consts");
diff --git a/src/xenia/gpu/spirv_shader_translator.h b/src/xenia/gpu/spirv_shader_translator.h
index 8138bbdc9..1ec006d50 100644
--- a/src/xenia/gpu/spirv_shader_translator.h
+++ b/src/xenia/gpu/spirv_shader_translator.h
@@ -22,6 +22,28 @@
namespace xe {
namespace gpu {
+// Push constants embedded within the command buffer.
+// The total size of this struct must be <= 128b (as that's the commonly
+// supported size).
+struct SpirvPushConstants {
+ // Accessible to vertex shader only:
+ float window_scale[4]; // sx,sy, ?, ?
+ float vtx_fmt[4];
+
+ // Accessible to fragment shader only:
+ float alpha_test[4]; // alpha test enable, func, ref, ?
+ uint32_t ps_param_gen;
+};
+static_assert(sizeof(SpirvPushConstants) <= 128,
+ "Push constants must fit <= 128b");
+constexpr uint32_t kSpirvPushConstantVertexRangeOffset = 0;
+constexpr uint32_t kSpirvPushConstantVertexRangeSize = (sizeof(float) * 4) * 2;
+constexpr uint32_t kSpirvPushConstantFragmentRangeOffset =
+ kSpirvPushConstantVertexRangeSize;
+constexpr uint32_t kSpirvPushConstantFragmentRangeSize =
+ (sizeof(float) * 4) + sizeof(uint32_t);
+constexpr uint32_t kSpirvPushConstantsSize = sizeof(SpirvPushConstants);
+
class SpirvShaderTranslator : public ShaderTranslator {
public:
SpirvShaderTranslator();
diff --git a/src/xenia/gpu/vulkan/pipeline_cache.cc b/src/xenia/gpu/vulkan/pipeline_cache.cc
index e86220f61..a8938e407 100644
--- a/src/xenia/gpu/vulkan/pipeline_cache.cc
+++ b/src/xenia/gpu/vulkan/pipeline_cache.cc
@@ -59,13 +59,12 @@ PipelineCache::PipelineCache(
// Push constants used for draw parameters.
// We need to keep these under 128b across all stages.
- VkPushConstantRange push_constant_ranges[2];
- push_constant_ranges[0].stageFlags = VK_SHADER_STAGE_VERTEX_BIT;
+ // TODO(benvanik): split between the stages?
+ VkPushConstantRange push_constant_ranges[1];
+ push_constant_ranges[0].stageFlags =
+ VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT;
push_constant_ranges[0].offset = 0;
- push_constant_ranges[0].size = sizeof(float) * 16;
- push_constant_ranges[1].stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
- push_constant_ranges[1].offset = sizeof(float) * 16;
- push_constant_ranges[1].size = sizeof(int);
+ push_constant_ranges[0].size = kSpirvPushConstantsSize;
// Shared pipeline layout.
VkPipelineLayoutCreateInfo pipeline_layout_info;
@@ -511,26 +510,74 @@ bool PipelineCache::SetDynamicState(VkCommandBuffer command_buffer,
// TODO(benvanik): push constants.
- bool push_constants_dirty = full_update;
+ bool push_constants_dirty = full_update || viewport_state_dirty;
push_constants_dirty |=
SetShadowRegister(®s.sq_program_cntl, XE_GPU_REG_SQ_PROGRAM_CNTL);
push_constants_dirty |=
SetShadowRegister(®s.sq_context_misc, XE_GPU_REG_SQ_CONTEXT_MISC);
+ push_constants_dirty |=
+ SetShadowRegister(®s.rb_colorcontrol, XE_GPU_REG_RB_COLORCONTROL);
+ push_constants_dirty |=
+ SetShadowRegister(®s.rb_alpha_ref, XE_GPU_REG_RB_ALPHA_REF);
+ if (push_constants_dirty) {
+ xenos::xe_gpu_program_cntl_t program_cntl;
+ program_cntl.dword_0 = regs.sq_program_cntl;
- xenos::xe_gpu_program_cntl_t program_cntl;
- program_cntl.dword_0 = regs.sq_program_cntl;
+ // Normal vertex shaders only, for now.
+ // TODO(benvanik): transform feedback/memexport.
+ // https://github.com/freedreno/freedreno/blob/master/includes/a2xx.xml.h
+ // 0 = normal
+ // 2 = point size
+ assert_true(program_cntl.vs_export_mode == 0 ||
+ program_cntl.vs_export_mode == 2);
- // Populate a register in the pixel shader with frag coord.
- int ps_param_gen = (regs.sq_context_misc >> 8) & 0xFF;
- // draw_batcher_.set_ps_param_gen(program_cntl.param_gen ? ps_param_gen : -1);
+ SpirvPushConstants push_constants;
- // Normal vertex shaders only, for now.
- // TODO(benvanik): transform feedback/memexport.
- // https://github.com/freedreno/freedreno/blob/master/includes/a2xx.xml.h
- // 0 = normal
- // 2 = point size
- assert_true(program_cntl.vs_export_mode == 0 ||
- program_cntl.vs_export_mode == 2);
+ // Done in VS, no need to flush state.
+ if ((regs.pa_cl_vte_cntl & (1 << 0)) > 0) {
+ push_constants.window_scale[0] = 1.0f;
+ push_constants.window_scale[1] = 1.0f;
+ } else {
+ push_constants.window_scale[0] = 1.0f / 2560.0f;
+ push_constants.window_scale[1] = -1.0f / 2560.0f;
+ }
+
+ // http://www.x.org/docs/AMD/old/evergreen_3D_registers_v2.pdf
+ // VTX_XY_FMT = true: the incoming XY have already been multiplied by 1/W0.
+ // = false: multiply the X, Y coordinates by 1/W0.
+ // VTX_Z_FMT = true: the incoming Z has already been multiplied by 1/W0.
+ // = false: multiply the Z coordinate by 1/W0.
+ // VTX_W0_FMT = true: the incoming W0 is not 1/W0. Perform the reciprocal to
+ // get 1/W0.
+ float vtx_xy_fmt = (regs.pa_cl_vte_cntl >> 8) & 0x1 ? 1.0f : 0.0f;
+ float vtx_z_fmt = (regs.pa_cl_vte_cntl >> 9) & 0x1 ? 1.0f : 0.0f;
+ float vtx_w0_fmt = (regs.pa_cl_vte_cntl >> 10) & 0x1 ? 1.0f : 0.0f;
+ push_constants.vtx_fmt[0] = vtx_xy_fmt;
+ push_constants.vtx_fmt[1] = vtx_xy_fmt;
+ push_constants.vtx_fmt[2] = vtx_z_fmt;
+ push_constants.vtx_fmt[3] = vtx_w0_fmt;
+
+ // Alpha testing -- ALPHAREF, ALPHAFUNC, ALPHATESTENABLE
+ // Deprecated in Vulkan, implemented in shader.
+ // if(ALPHATESTENABLE && frag_out.a [<=/ALPHAFUNC] ALPHAREF) discard;
+ // ALPHATESTENABLE
+ push_constants.alpha_test[0] =
+ (regs.rb_colorcontrol & 0x8) != 0 ? 1.0f : 0.0f;
+ // ALPHAFUNC
+ push_constants.alpha_test[1] =
+ static_cast<float>(regs.rb_colorcontrol & 0x7);
+ // ALPHAREF
+ push_constants.alpha_test[2] = regs.rb_alpha_ref;
+
+ // Whether to populate a register in the pixel shader with frag coord.
+ int ps_param_gen = (regs.sq_context_misc >> 8) & 0xFF;
+ push_constants.ps_param_gen = program_cntl.param_gen ? ps_param_gen : -1;
+
+ vkCmdPushConstants(
+ command_buffer, pipeline_layout_,
+ VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT, 0,
+ kSpirvPushConstantsSize, &push_constants);
+ }
return true;
}
diff --git a/src/xenia/gpu/vulkan/pipeline_cache.h b/src/xenia/gpu/vulkan/pipeline_cache.h
index 7d35fc496..3e623f14e 100644
--- a/src/xenia/gpu/vulkan/pipeline_cache.h
+++ b/src/xenia/gpu/vulkan/pipeline_cache.h
@@ -262,6 +262,8 @@ class PipelineCache {
uint32_t sq_program_cntl;
uint32_t sq_context_misc;
+ uint32_t rb_colorcontrol;
+ float rb_alpha_ref;
SetDynamicStateRegisters() { Reset(); }
void Reset() { std::memset(this, 0, sizeof(*this)); }