#include "nir/nir_deref.h"
#include "spirv_info.h"
+#include "util/format/u_format.h"
#include "util/u_math.h"
#include <stdio.h>
longjmp(b->fail_jump, 1);
}
-struct spec_constant_value {
- bool is_double;
- union {
- uint32_t data32;
- uint64_t data64;
- };
-};
-
static struct vtn_ssa_value *
vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
{
return w;
}
+static bool
+vtn_handle_non_semantic_instruction(struct vtn_builder *b, SpvOp ext_opcode,
+ const uint32_t *w, unsigned count)
+{
+ /* Do nothing. */
+ return true;
+}
+
static void
vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
} else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0)
&& (b->options && b->options->caps.amd_trinary_minmax)) {
val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
+ } else if ((strcmp(ext, "SPV_AMD_shader_explicit_vertex_parameter") == 0)
+ && (b->options && b->options->caps.amd_shader_explicit_vertex_parameter)) {
+ val->ext_handler = vtn_handle_amd_shader_explicit_vertex_parameter_instruction;
} else if (strcmp(ext, "OpenCL.std") == 0) {
val->ext_handler = vtn_handle_opencl_instruction;
+ } else if (strstr(ext, "NonSemantic.") == ext) {
+ val->ext_handler = vtn_handle_non_semantic_instruction;
} else {
vtn_fail("Unsupported extension: %s", ext);
}
static void
struct_member_decoration_cb(struct vtn_builder *b,
- struct vtn_value *val, int member,
+ UNUSED struct vtn_value *val, int member,
const struct vtn_decoration *dec, void *void_ctx)
{
struct member_decoration_ctx *ctx = void_ctx;
case SpvDecorationFlat:
ctx->fields[member].interpolation = INTERP_MODE_FLAT;
break;
+ case SpvDecorationExplicitInterpAMD:
+ ctx->fields[member].interpolation = INTERP_MODE_EXPLICIT;
+ break;
case SpvDecorationCentroid:
ctx->fields[member].centroid = true;
break;
ctx->fields[member].sample = true;
break;
case SpvDecorationStream:
- /* Vulkan only allows one GS stream */
- vtn_assert(dec->operands[0] == 0);
+ /* This is handled later by var_decoration_cb in vtn_variables.c */
break;
case SpvDecorationLocation:
ctx->fields[member].location = dec->operands[0];
case SpvDecorationXfbBuffer:
case SpvDecorationXfbStride:
- vtn_warn("Vulkan does not have transform feedback");
+ /* This is handled later by var_decoration_cb in vtn_variables.c */
break;
case SpvDecorationCPacked:
*/
static void
struct_member_matrix_stride_cb(struct vtn_builder *b,
- struct vtn_value *val, int member,
+ UNUSED struct vtn_value *val, int member,
const struct vtn_decoration *dec,
void *void_ctx)
{
static void
type_decoration_cb(struct vtn_builder *b,
struct vtn_value *val, int member,
- const struct vtn_decoration *dec, void *ctx)
+ const struct vtn_decoration *dec, UNUSED void *ctx)
{
struct vtn_type *type = val->type;
case SpvDecorationPatch:
case SpvDecorationCentroid:
case SpvDecorationSample:
+ case SpvDecorationExplicitInterpAMD:
case SpvDecorationVolatile:
case SpvDecorationCoherent:
case SpvDecorationNonWritable:
translate_image_format(struct vtn_builder *b, SpvImageFormat format)
{
switch (format) {
- case SpvImageFormatUnknown: return 0; /* GL_NONE */
- case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */
- case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */
- case SpvImageFormatR32f: return 0x822E; /* GL_R32F */
- case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */
- case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */
- case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */
- case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */
- case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
- case SpvImageFormatR16f: return 0x822D; /* GL_R16F */
- case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */
- case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */
- case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */
- case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */
- case SpvImageFormatR16: return 0x822A; /* GL_R16 */
- case SpvImageFormatR8: return 0x8229; /* GL_R8 */
- case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */
- case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */
- case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */
- case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */
- case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */
- case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */
- case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */
- case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */
- case SpvImageFormatR32i: return 0x8235; /* GL_R32I */
- case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */
- case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */
- case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */
- case SpvImageFormatR16i: return 0x8233; /* GL_R16I */
- case SpvImageFormatR8i: return 0x8231; /* GL_R8I */
- case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */
- case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */
- case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */
- case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */
- case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */
- case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */
- case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */
- case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */
- case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */
- case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */
+ case SpvImageFormatUnknown: return PIPE_FORMAT_NONE;
+ case SpvImageFormatRgba32f: return PIPE_FORMAT_R32G32B32A32_FLOAT;
+ case SpvImageFormatRgba16f: return PIPE_FORMAT_R16G16B16A16_FLOAT;
+ case SpvImageFormatR32f: return PIPE_FORMAT_R32_FLOAT;
+ case SpvImageFormatRgba8: return PIPE_FORMAT_R8G8B8A8_UNORM;
+ case SpvImageFormatRgba8Snorm: return PIPE_FORMAT_R8G8B8A8_SNORM;
+ case SpvImageFormatRg32f: return PIPE_FORMAT_R32G32_FLOAT;
+ case SpvImageFormatRg16f: return PIPE_FORMAT_R16G16_FLOAT;
+ case SpvImageFormatR11fG11fB10f: return PIPE_FORMAT_R11G11B10_FLOAT;
+ case SpvImageFormatR16f: return PIPE_FORMAT_R16_FLOAT;
+ case SpvImageFormatRgba16: return PIPE_FORMAT_R16G16B16A16_UNORM;
+ case SpvImageFormatRgb10A2: return PIPE_FORMAT_R10G10B10A2_UNORM;
+ case SpvImageFormatRg16: return PIPE_FORMAT_R16G16_UNORM;
+ case SpvImageFormatRg8: return PIPE_FORMAT_R8G8_UNORM;
+ case SpvImageFormatR16: return PIPE_FORMAT_R16_UNORM;
+ case SpvImageFormatR8: return PIPE_FORMAT_R8_UNORM;
+ case SpvImageFormatRgba16Snorm: return PIPE_FORMAT_R16G16B16A16_SNORM;
+ case SpvImageFormatRg16Snorm: return PIPE_FORMAT_R16G16_SNORM;
+ case SpvImageFormatRg8Snorm: return PIPE_FORMAT_R8G8_SNORM;
+ case SpvImageFormatR16Snorm: return PIPE_FORMAT_R16_SNORM;
+ case SpvImageFormatR8Snorm: return PIPE_FORMAT_R8_SNORM;
+ case SpvImageFormatRgba32i: return PIPE_FORMAT_R32G32B32A32_SINT;
+ case SpvImageFormatRgba16i: return PIPE_FORMAT_R16G16B16A16_SINT;
+ case SpvImageFormatRgba8i: return PIPE_FORMAT_R8G8B8A8_SINT;
+ case SpvImageFormatR32i: return PIPE_FORMAT_R32_SINT;
+ case SpvImageFormatRg32i: return PIPE_FORMAT_R32G32_SINT;
+ case SpvImageFormatRg16i: return PIPE_FORMAT_R16G16_SINT;
+ case SpvImageFormatRg8i: return PIPE_FORMAT_R8G8_SINT;
+ case SpvImageFormatR16i: return PIPE_FORMAT_R16_SINT;
+ case SpvImageFormatR8i: return PIPE_FORMAT_R8_SINT;
+ case SpvImageFormatRgba32ui: return PIPE_FORMAT_R32G32B32A32_UINT;
+ case SpvImageFormatRgba16ui: return PIPE_FORMAT_R16G16B16A16_UINT;
+ case SpvImageFormatRgba8ui: return PIPE_FORMAT_R8G8B8A8_UINT;
+ case SpvImageFormatR32ui: return PIPE_FORMAT_R32_UINT;
+ case SpvImageFormatRgb10a2ui: return PIPE_FORMAT_R10G10B10A2_UINT;
+ case SpvImageFormatRg32ui: return PIPE_FORMAT_R32G32_UINT;
+ case SpvImageFormatRg16ui: return PIPE_FORMAT_R16G16_UINT;
+ case SpvImageFormatRg8ui: return PIPE_FORMAT_R8G8_UINT;
+ case SpvImageFormatR16ui: return PIPE_FORMAT_R16_UINT;
+ case SpvImageFormatR8ui: return PIPE_FORMAT_R8_UINT;
default:
vtn_fail("Invalid image format: %s (%u)",
spirv_imageformat_to_string(format), format);
case SpvStorageClassUniform:
case SpvStorageClassPushConstant:
case SpvStorageClassStorageBuffer:
- case SpvStorageClassPhysicalStorageBufferEXT:
+ case SpvStorageClassPhysicalStorageBuffer:
vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
break;
default:
case SpvStorageClassFunction:
case SpvStorageClassWorkgroup:
case SpvStorageClassCrossWorkgroup:
+ case SpvStorageClassUniformConstant:
val->type->stride = align(glsl_get_cl_size(val->type->deref->type),
glsl_get_cl_alignment(val->type->deref->type));
break;
}
static void
-spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v,
- int member, const struct vtn_decoration *dec,
- void *data)
+spec_constant_decoration_cb(struct vtn_builder *b, UNUSED struct vtn_value *val,
+ ASSERTED int member,
+ const struct vtn_decoration *dec, void *data)
{
vtn_assert(member == -1);
if (dec->decoration != SpvDecorationSpecId)
return;
- struct spec_constant_value *const_value = data;
-
+ nir_const_value *value = data;
for (unsigned i = 0; i < b->num_specializations; i++) {
if (b->specializations[i].id == dec->operands[0]) {
- if (const_value->is_double)
- const_value->data64 = b->specializations[i].data64;
- else
- const_value->data32 = b->specializations[i].data32;
+ *value = b->specializations[i].value;
return;
}
}
}
-static uint32_t
-get_specialization(struct vtn_builder *b, struct vtn_value *val,
- uint32_t const_value)
-{
- struct spec_constant_value data;
- data.is_double = false;
- data.data32 = const_value;
- vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
- return data.data32;
-}
-
-static uint64_t
-get_specialization64(struct vtn_builder *b, struct vtn_value *val,
- uint64_t const_value)
-{
- struct spec_constant_value data;
- data.is_double = true;
- data.data64 = const_value;
- vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data);
- return data.data64;
-}
-
static void
handle_workgroup_size_decoration_cb(struct vtn_builder *b,
struct vtn_value *val,
- int member,
+ ASSERTED int member,
const struct vtn_decoration *dec,
- void *data)
+ UNUSED void *data)
{
vtn_assert(member == -1);
if (dec->decoration != SpvDecorationBuiltIn ||
"Result type of %s must be OpTypeBool",
spirv_op_to_string(opcode));
- uint32_t int_val = (opcode == SpvOpConstantTrue ||
- opcode == SpvOpSpecConstantTrue);
+ bool bval = (opcode == SpvOpConstantTrue ||
+ opcode == SpvOpSpecConstantTrue);
+
+ nir_const_value u32val = nir_const_value_for_uint(bval, 32);
if (opcode == SpvOpSpecConstantTrue ||
opcode == SpvOpSpecConstantFalse)
- int_val = get_specialization(b, val, int_val);
+ vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &u32val);
- val->constant->values[0].b = int_val != 0;
+ val->constant->values[0].b = u32val.u32 != 0;
break;
}
- case SpvOpConstant: {
+ case SpvOpConstant:
+ case SpvOpSpecConstant: {
vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
"Result type of %s must be a scalar",
spirv_op_to_string(opcode));
default:
vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size);
}
- break;
- }
- case SpvOpSpecConstant: {
- vtn_fail_if(val->type->base_type != vtn_base_type_scalar,
- "Result type of %s must be a scalar",
- spirv_op_to_string(opcode));
- int bit_size = glsl_get_bit_size(val->type->type);
- switch (bit_size) {
- case 64:
- val->constant->values[0].u64 =
- get_specialization64(b, val, vtn_u64_literal(&w[3]));
- break;
- case 32:
- val->constant->values[0].u32 = get_specialization(b, val, w[3]);
- break;
- case 16:
- val->constant->values[0].u16 = get_specialization(b, val, w[3]);
- break;
- case 8:
- val->constant->values[0].u8 = get_specialization(b, val, w[3]);
- break;
- default:
- vtn_fail("Unsupported SpvOpSpecConstant bit size");
- }
+ if (opcode == SpvOpSpecConstant)
+ vtn_foreach_decoration(b, val, spec_constant_decoration_cb,
+ &val->constant->values[0]);
break;
}
}
case SpvOpSpecConstantOp: {
- SpvOp opcode = get_specialization(b, val, w[3]);
+ nir_const_value u32op = nir_const_value_for_uint(w[3], 32);
+ vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &u32op);
+ SpvOp opcode = u32op.u32;
switch (opcode) {
case SpvOpVectorShuffle: {
struct vtn_value *v0 = &b->values[w[4]];
{
switch (sc) {
case SpvStorageClassStorageBuffer:
- case SpvStorageClassPhysicalStorageBufferEXT:
+ case SpvStorageClassPhysicalStorageBuffer:
return SpvMemorySemanticsUniformMemoryMask;
case SpvStorageClassWorkgroup:
return SpvMemorySemanticsWorkgroupMemoryMask;
}
}
+static void
+vtn_split_barrier_semantics(struct vtn_builder *b,
+ SpvMemorySemanticsMask semantics,
+ SpvMemorySemanticsMask *before,
+ SpvMemorySemanticsMask *after)
+{
+ /* For memory semantics embedded in operations, we split them into up to
+ * two barriers, to be added before and after the operation. This is less
+ * strict than if we propagated until the final backend stage, but still
+ * result in correct execution.
+ *
+ * A further improvement could be pipe this information (and use!) into the
+ * next compiler layers, at the expense of making the handling of barriers
+ * more complicated.
+ */
+
+ *before = SpvMemorySemanticsMaskNone;
+ *after = SpvMemorySemanticsMaskNone;
+
+ SpvMemorySemanticsMask order_semantics =
+ semantics & (SpvMemorySemanticsAcquireMask |
+ SpvMemorySemanticsReleaseMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask);
+
+ if (util_bitcount(order_semantics) > 1) {
+ /* Old GLSLang versions incorrectly set all the ordering bits. This was
+ * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+ * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+ */
+ vtn_warn("Multiple memory ordering semantics specified, "
+ "assuming AcquireRelease.");
+ order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+ }
+
+ const SpvMemorySemanticsMask av_vis_semantics =
+ semantics & (SpvMemorySemanticsMakeAvailableMask |
+ SpvMemorySemanticsMakeVisibleMask);
+
+ const SpvMemorySemanticsMask storage_semantics =
+ semantics & (SpvMemorySemanticsUniformMemoryMask |
+ SpvMemorySemanticsSubgroupMemoryMask |
+ SpvMemorySemanticsWorkgroupMemoryMask |
+ SpvMemorySemanticsCrossWorkgroupMemoryMask |
+ SpvMemorySemanticsAtomicCounterMemoryMask |
+ SpvMemorySemanticsImageMemoryMask |
+ SpvMemorySemanticsOutputMemoryMask);
+
+ const SpvMemorySemanticsMask other_semantics =
+ semantics & ~(order_semantics | av_vis_semantics | storage_semantics);
+
+ if (other_semantics)
+ vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics);
+
+ /* SequentiallyConsistent is treated as AcquireRelease. */
+
+ /* The RELEASE barrier happens BEFORE the operation, and it is usually
+ * associated with a Store. All the write operations with a matching
+ * semantics will not be reordered after the Store.
+ */
+ if (order_semantics & (SpvMemorySemanticsReleaseMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask)) {
+ *before |= SpvMemorySemanticsReleaseMask | storage_semantics;
+ }
+
+ /* The ACQUIRE barrier happens AFTER the operation, and it is usually
+ * associated with a Load. All the operations with a matching semantics
+ * will not be reordered before the Load.
+ */
+ if (order_semantics & (SpvMemorySemanticsAcquireMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask)) {
+ *after |= SpvMemorySemanticsAcquireMask | storage_semantics;
+ }
+
+ if (av_vis_semantics & SpvMemorySemanticsMakeVisibleMask)
+ *before |= SpvMemorySemanticsMakeVisibleMask | storage_semantics;
+
+ if (av_vis_semantics & SpvMemorySemanticsMakeAvailableMask)
+ *after |= SpvMemorySemanticsMakeAvailableMask | storage_semantics;
+}
+
+static void
+vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope,
+ SpvMemorySemanticsMask semantics)
+{
+ nir_memory_semantics nir_semantics = 0;
+
+ SpvMemorySemanticsMask order_semantics =
+ semantics & (SpvMemorySemanticsAcquireMask |
+ SpvMemorySemanticsReleaseMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask);
+
+ if (util_bitcount(order_semantics) > 1) {
+ /* Old GLSLang versions incorrectly set all the ordering bits. This was
+ * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+ * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+ */
+ vtn_warn("Multiple memory ordering semantics bits specified, "
+ "assuming AcquireRelease.");
+ order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+ }
+
+ switch (order_semantics) {
+ case 0:
+ /* Not an ordering barrier. */
+ break;
+
+ case SpvMemorySemanticsAcquireMask:
+ nir_semantics = NIR_MEMORY_ACQUIRE;
+ break;
+
+ case SpvMemorySemanticsReleaseMask:
+ nir_semantics = NIR_MEMORY_RELEASE;
+ break;
+
+ case SpvMemorySemanticsSequentiallyConsistentMask:
+ /* Fall through. Treated as AcquireRelease in Vulkan. */
+ case SpvMemorySemanticsAcquireReleaseMask:
+ nir_semantics = NIR_MEMORY_ACQUIRE | NIR_MEMORY_RELEASE;
+ break;
+
+ default:
+ unreachable("Invalid memory order semantics");
+ }
+
+ if (semantics & SpvMemorySemanticsMakeAvailableMask) {
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use MakeAvailable memory semantics the VulkanMemoryModel "
+ "capability must be declared.");
+ nir_semantics |= NIR_MEMORY_MAKE_AVAILABLE;
+ }
+
+ if (semantics & SpvMemorySemanticsMakeVisibleMask) {
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use MakeVisible memory semantics the VulkanMemoryModel "
+ "capability must be declared.");
+ nir_semantics |= NIR_MEMORY_MAKE_VISIBLE;
+ }
+
+ /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory,
+ * and AtomicCounterMemory are ignored".
+ */
+ semantics &= ~(SpvMemorySemanticsSubgroupMemoryMask |
+ SpvMemorySemanticsCrossWorkgroupMemoryMask |
+ SpvMemorySemanticsAtomicCounterMemoryMask);
+
+ /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used
+ * for SpvMemorySemanticsImageMemoryMask.
+ */
+
+ nir_variable_mode modes = 0;
+ if (semantics & (SpvMemorySemanticsUniformMemoryMask |
+ SpvMemorySemanticsImageMemoryMask)) {
+ modes |= nir_var_uniform |
+ nir_var_mem_ubo |
+ nir_var_mem_ssbo |
+ nir_var_mem_global;
+ }
+ if (semantics & SpvMemorySemanticsWorkgroupMemoryMask)
+ modes |= nir_var_mem_shared;
+ if (semantics & SpvMemorySemanticsOutputMemoryMask) {
+ modes |= nir_var_shader_out;
+ }
+
+ /* No barrier to add. */
+ if (nir_semantics == 0 || modes == 0)
+ return;
+
+ nir_scope nir_scope;
+ switch (scope) {
+ case SpvScopeDevice:
+ vtn_fail_if(b->options->caps.vk_memory_model &&
+ !b->options->caps.vk_memory_model_device_scope,
+ "If the Vulkan memory model is declared and any instruction "
+ "uses Device scope, the VulkanMemoryModelDeviceScope "
+ "capability must be declared.");
+ nir_scope = NIR_SCOPE_DEVICE;
+ break;
+
+ case SpvScopeQueueFamily:
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "To use Queue Family scope, the VulkanMemoryModel capability "
+ "must be declared.");
+ nir_scope = NIR_SCOPE_QUEUE_FAMILY;
+ break;
+
+ case SpvScopeWorkgroup:
+ nir_scope = NIR_SCOPE_WORKGROUP;
+ break;
+
+ case SpvScopeSubgroup:
+ nir_scope = NIR_SCOPE_SUBGROUP;
+ break;
+
+ case SpvScopeInvocation:
+ nir_scope = NIR_SCOPE_INVOCATION;
+ break;
+
+ default:
+ vtn_fail("Invalid memory scope");
+ }
+
+ nir_intrinsic_instr *intrin =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_scoped_memory_barrier);
+ nir_intrinsic_set_memory_semantics(intrin, nir_semantics);
+
+ nir_intrinsic_set_memory_modes(intrin, modes);
+ nir_intrinsic_set_memory_scope(intrin, nir_scope);
+ nir_builder_instr_insert(&b->nb, &intrin->instr);
+}
+
struct vtn_ssa_value *
vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
{
return src;
}
+static uint32_t
+image_operand_arg(struct vtn_builder *b, const uint32_t *w, uint32_t count,
+ uint32_t mask_idx, SpvImageOperandsMask op)
+{
+ static const SpvImageOperandsMask ops_with_arg =
+ SpvImageOperandsBiasMask |
+ SpvImageOperandsLodMask |
+ SpvImageOperandsGradMask |
+ SpvImageOperandsConstOffsetMask |
+ SpvImageOperandsOffsetMask |
+ SpvImageOperandsConstOffsetsMask |
+ SpvImageOperandsSampleMask |
+ SpvImageOperandsMinLodMask |
+ SpvImageOperandsMakeTexelAvailableMask |
+ SpvImageOperandsMakeTexelVisibleMask;
+
+ assert(util_bitcount(op) == 1);
+ assert(w[mask_idx] & op);
+ assert(op & ops_with_arg);
+
+ uint32_t idx = util_bitcount(w[mask_idx] & (op - 1) & ops_with_arg) + 1;
+
+ /* Adjust indices for operands with two arguments. */
+ static const SpvImageOperandsMask ops_with_two_args =
+ SpvImageOperandsGradMask;
+ idx += util_bitcount(w[mask_idx] & (op - 1) & ops_with_two_args);
+
+ idx += mask_idx;
+
+ vtn_fail_if(idx + (op & ops_with_two_args ? 1 : 0) >= count,
+ "Image op claims to have %s but does not enough "
+ "following operands", spirv_imageoperands_to_string(op));
+
+ return idx;
+}
+
static void
vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
struct vtn_value *val =
vtn_push_value(b, w[2], vtn_value_type_sampled_image);
val->sampled_image = ralloc(b, struct vtn_sampled_image);
- val->sampled_image->type =
- vtn_value(b, w[1], vtn_value_type_type)->type;
val->sampled_image->image =
vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
val->sampled_image->sampler =
struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
- struct vtn_sampled_image sampled;
+ struct vtn_pointer *image = NULL, *sampler = NULL;
struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
if (sampled_val->value_type == vtn_value_type_sampled_image) {
- sampled = *sampled_val->sampled_image;
+ image = sampled_val->sampled_image->image;
+ sampler = sampled_val->sampled_image->sampler;
} else {
vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
- sampled.type = sampled_val->pointer->type;
- sampled.image = NULL;
- sampled.sampler = sampled_val->pointer;
+ image = sampled_val->pointer;
}
- const struct glsl_type *image_type = sampled.type->type;
+ nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image);
+ nir_deref_instr *sampler_deref =
+ sampler ? vtn_pointer_to_deref(b, sampler) : NULL;
+
+ const struct glsl_type *image_type = sampled_val->type->type;
const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
const bool is_array = glsl_sampler_type_is_array(image_type);
nir_alu_type dest_type = nir_type_invalid;
break;
case SpvOpImageFetch:
- if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
+ if (sampler_dim == GLSL_SAMPLER_DIM_MS) {
texop = nir_texop_txf_ms;
} else {
texop = nir_texop_txf;
dest_type = nir_type_int;
break;
+ case SpvOpFragmentFetchAMD:
+ texop = nir_texop_fragment_fetch;
+ break;
+
+ case SpvOpFragmentMaskFetchAMD:
+ texop = nir_texop_fragment_mask_fetch;
+ break;
+
default:
vtn_fail_with_opcode("Unhandled opcode", opcode);
}
nir_tex_src srcs[10]; /* 10 should be enough */
nir_tex_src *p = srcs;
- nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
- nir_deref_instr *texture =
- sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
-
- p->src = nir_src_for_ssa(&texture->dest.ssa);
+ p->src = nir_src_for_ssa(&image_deref->dest.ssa);
p->src_type = nir_tex_src_texture_deref;
p++;
case nir_texop_txd:
case nir_texop_tg4:
case nir_texop_lod:
- /* These operations require a sampler */
- p->src = nir_src_for_ssa(&sampler->dest.ssa);
+ vtn_fail_if(sampler == NULL,
+ "%s requires an image of type OpTypeSampledImage",
+ spirv_op_to_string(opcode));
+ p->src = nir_src_for_ssa(&sampler_deref->dest.ssa);
p->src_type = nir_tex_src_sampler_deref;
p++;
break;
case nir_texop_query_levels:
case nir_texop_texture_samples:
case nir_texop_samples_identical:
+ case nir_texop_fragment_fetch:
+ case nir_texop_fragment_mask_fetch:
/* These don't */
break;
case nir_texop_txf_ms_fb:
case SpvOpImageFetch:
case SpvOpImageGather:
case SpvOpImageDrefGather:
- case SpvOpImageQueryLod: {
+ case SpvOpImageQueryLod:
+ case SpvOpFragmentFetchAMD:
+ case SpvOpFragmentMaskFetchAMD: {
/* All these types have the coordinate as their first real argument */
- switch (sampler_dim) {
- case GLSL_SAMPLER_DIM_1D:
- case GLSL_SAMPLER_DIM_BUF:
- coord_components = 1;
- break;
- case GLSL_SAMPLER_DIM_2D:
- case GLSL_SAMPLER_DIM_RECT:
- case GLSL_SAMPLER_DIM_MS:
- coord_components = 2;
- break;
- case GLSL_SAMPLER_DIM_3D:
- case GLSL_SAMPLER_DIM_CUBE:
- coord_components = 3;
- break;
- default:
- vtn_fail("Invalid sampler type");
- }
+ coord_components = glsl_get_sampler_dim_coordinate_components(sampler_dim);
if (is_array && texop != nir_texop_lod)
coord_components++;
if (opcode == SpvOpImageQuerySizeLod)
(*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+ /* For OpFragmentFetchAMD, we always have a multisample index */
+ if (opcode == SpvOpFragmentFetchAMD)
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
+
/* Now we need to handle some number of optional arguments */
struct vtn_value *gather_offsets = NULL;
if (idx < count) {
- uint32_t operands = w[idx++];
+ uint32_t operands = w[idx];
if (operands & SpvImageOperandsBiasMask) {
vtn_assert(texop == nir_texop_tex);
texop = nir_texop_txb;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsBiasMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_bias);
}
if (operands & SpvImageOperandsLodMask) {
vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
texop == nir_texop_txs);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsLodMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_lod);
}
if (operands & SpvImageOperandsGradMask) {
vtn_assert(texop == nir_texop_txl);
texop = nir_texop_txd;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsGradMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ddx);
+ (*p++) = vtn_tex_src(b, w[arg + 1], nir_tex_src_ddy);
+ }
+
+ vtn_fail_if(util_bitcount(operands & (SpvImageOperandsConstOffsetsMask |
+ SpvImageOperandsOffsetMask |
+ SpvImageOperandsConstOffsetMask)) > 1,
+ "At most one of the ConstOffset, Offset, and ConstOffsets "
+ "image operands can be used on a given instruction.");
+
+ if (operands & SpvImageOperandsOffsetMask) {
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsOffsetMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
}
- if (operands & SpvImageOperandsOffsetMask ||
- operands & SpvImageOperandsConstOffsetMask)
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
+ if (operands & SpvImageOperandsConstOffsetMask) {
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsConstOffsetMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
+ }
if (operands & SpvImageOperandsConstOffsetsMask) {
vtn_assert(texop == nir_texop_tg4);
- gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsConstOffsetsMask);
+ gather_offsets = vtn_value(b, w[arg], vtn_value_type_constant);
}
if (operands & SpvImageOperandsSampleMask) {
vtn_assert(texop == nir_texop_txf_ms);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsSampleMask);
texop = nir_texop_txf_ms;
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ms_index);
}
if (operands & SpvImageOperandsMinLodMask) {
vtn_assert(texop == nir_texop_tex ||
texop == nir_texop_txb ||
texop == nir_texop_txd);
- (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod);
+ uint32_t arg = image_operand_arg(b, w, count, idx,
+ SpvImageOperandsMinLodMask);
+ (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_min_lod);
}
}
- /* We should have now consumed exactly all of the arguments */
- vtn_assert(idx == count);
nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
instr->op = texop;
is_shadow && glsl_get_components(ret_type->type) == 1;
instr->component = gather_component;
- if (sampled.image && (sampled.image->access & ACCESS_NON_UNIFORM))
+ if (image && (image->access & ACCESS_NON_UNIFORM))
instr->texture_non_uniform = true;
- if (sampled.sampler && (sampled.sampler->access & ACCESS_NON_UNIFORM))
+ if (sampler && (sampler->access & ACCESS_NON_UNIFORM))
instr->sampler_non_uniform = true;
/* for non-query ops, get dest_type from sampler type */
val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
val->image->coord = get_image_coord(b, w[4]);
val->image->sample = vtn_ssa_value(b, w[5])->def;
+ val->image->lod = nir_imm_int(&b->nb, 0);
return;
}
image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
image.coord = NULL;
image.sample = NULL;
+ image.lod = NULL;
break;
- case SpvOpImageRead:
+ case SpvOpImageRead: {
image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
image.coord = get_image_coord(b, w[4]);
- if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
- vtn_assert(w[5] == SpvImageOperandsSampleMask);
- image.sample = vtn_ssa_value(b, w[6])->def;
+ const SpvImageOperandsMask operands =
+ count > 5 ? w[5] : SpvImageOperandsMaskNone;
+
+ if (operands & SpvImageOperandsSampleMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 5,
+ SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[arg])->def;
} else {
image.sample = nir_ssa_undef(&b->nb, 1, 32);
}
+
+ if (operands & SpvImageOperandsMakeTexelVisibleMask) {
+ vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
+ "MakeTexelVisible requires NonPrivateTexel to also be set.");
+ uint32_t arg = image_operand_arg(b, w, count, 5,
+ SpvImageOperandsMakeTexelVisibleMask);
+ semantics = SpvMemorySemanticsMakeVisibleMask;
+ scope = vtn_constant_uint(b, w[arg]);
+ }
+
+ if (operands & SpvImageOperandsLodMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 5,
+ SpvImageOperandsLodMask);
+ image.lod = vtn_ssa_value(b, w[arg])->def;
+ } else {
+ image.lod = nir_imm_int(&b->nb, 0);
+ }
+
+ /* TODO: Volatile. */
+
break;
+ }
- case SpvOpImageWrite:
+ case SpvOpImageWrite: {
image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer;
image.coord = get_image_coord(b, w[2]);
/* texel = w[3] */
- if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
- vtn_assert(w[4] == SpvImageOperandsSampleMask);
- image.sample = vtn_ssa_value(b, w[5])->def;
+ const SpvImageOperandsMask operands =
+ count > 4 ? w[4] : SpvImageOperandsMaskNone;
+
+ if (operands & SpvImageOperandsSampleMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 4,
+ SpvImageOperandsSampleMask);
+ image.sample = vtn_ssa_value(b, w[arg])->def;
} else {
image.sample = nir_ssa_undef(&b->nb, 1, 32);
}
+
+ if (operands & SpvImageOperandsMakeTexelAvailableMask) {
+ vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
+ "MakeTexelAvailable requires NonPrivateTexel to also be set.");
+ uint32_t arg = image_operand_arg(b, w, count, 4,
+ SpvImageOperandsMakeTexelAvailableMask);
+ semantics = SpvMemorySemanticsMakeAvailableMask;
+ scope = vtn_constant_uint(b, w[arg]);
+ }
+
+ if (operands & SpvImageOperandsLodMask) {
+ uint32_t arg = image_operand_arg(b, w, count, 4,
+ SpvImageOperandsLodMask);
+ image.lod = vtn_ssa_value(b, w[arg])->def;
+ } else {
+ image.lod = nir_imm_int(&b->nb, 0);
+ }
+
+ /* TODO: Volatile. */
+
break;
+ }
default:
vtn_fail_with_opcode("Invalid image opcode", opcode);
case SpvOpAtomicLoad:
case SpvOpImageQuerySize:
case SpvOpImageRead:
+ if (opcode == SpvOpImageRead || opcode == SpvOpAtomicLoad) {
+ /* Only OpImageRead can support a lod parameter if
+ * SPV_AMD_shader_image_load_store_lod is used but the current NIR
+ * intrinsics definition for atomics requires us to set it for
+ * OpAtomicLoad.
+ */
+ intrin->src[3] = nir_src_for_ssa(image.lod);
+ }
break;
case SpvOpAtomicStore:
case SpvOpImageWrite: {
assert(op == nir_intrinsic_image_deref_store);
intrin->num_components = 4;
intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value));
+ /* Only OpImageWrite can support a lod parameter if
+ * SPV_AMD_shader_image_load_store_lod is used but the current NIR
+ * intrinsics definition for atomics requires us to set it for
+ * OpAtomicStore.
+ */
+ intrin->src[4] = nir_src_for_ssa(image.lod);
break;
}
/* Image operations implicitly have the Image storage memory semantics. */
semantics |= SpvMemorySemanticsImageMemoryMask;
+ SpvMemorySemanticsMask before_semantics;
+ SpvMemorySemanticsMask after_semantics;
+ vtn_split_barrier_semantics(b, semantics, &before_semantics, &after_semantics);
+
+ if (before_semantics)
+ vtn_emit_memory_barrier(b, scope, before_semantics);
+
if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) {
struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
} else {
nir_builder_instr_insert(&b->nb, &intrin->instr);
}
+
+ if (after_semantics)
+ vtn_emit_memory_barrier(b, scope, after_semantics);
}
static nir_intrinsic_op
* only need to support GLSL Atomic Counters that are uints and don't
* allow direct storage.
*/
- unreachable("Invalid uniform atomic");
+ vtn_fail("Invalid uniform atomic");
}
}
*/
static void
vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
- const uint32_t *w, unsigned count)
+ const uint32_t *w, UNUSED unsigned count)
{
struct vtn_pointer *ptr;
nir_intrinsic_instr *atomic;
*/
semantics |= vtn_storage_class_to_memory_semantics(ptr->ptr_type->storage_class);
+ SpvMemorySemanticsMask before_semantics;
+ SpvMemorySemanticsMask after_semantics;
+ vtn_split_barrier_semantics(b, semantics, &before_semantics, &after_semantics);
+
+ if (before_semantics)
+ vtn_emit_memory_barrier(b, scope, before_semantics);
+
if (opcode != SpvOpAtomicStore) {
struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
}
nir_builder_instr_insert(&b->nb, &atomic->instr);
+
+ if (after_semantics)
+ vtn_emit_memory_barrier(b, scope, after_semantics);
}
static nir_alu_instr *
return dest;
}
-nir_ssa_def *
-vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
-{
- return nir_channel(&b->nb, src, index);
-}
-
-nir_ssa_def *
-vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
- unsigned index)
-{
- nir_alu_instr *vec = create_vec(b, src->num_components,
- src->bit_size);
-
- for (unsigned i = 0; i < src->num_components; i++) {
- if (i == index) {
- vec->src[i].src = nir_src_for_ssa(insert);
- } else {
- vec->src[i].src = nir_src_for_ssa(src);
- vec->src[i].swizzle[0] = i;
- }
- }
-
- nir_builder_instr_insert(&b->nb, &vec->instr);
-
- return &vec->dest.dest.ssa;
-}
-
-static nir_ssa_def *
-nir_ieq_imm(nir_builder *b, nir_ssa_def *x, uint64_t i)
-{
- return nir_ieq(b, x, nir_imm_intN_t(b, i, x->bit_size));
-}
-
-nir_ssa_def *
-vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
- nir_ssa_def *index)
-{
- return nir_vector_extract(&b->nb, src, nir_i2i(&b->nb, index, 32));
-}
-
-nir_ssa_def *
-vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
- nir_ssa_def *insert, nir_ssa_def *index)
-{
- nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
- for (unsigned i = 1; i < src->num_components; i++)
- dest = nir_bcsel(&b->nb, nir_ieq_imm(&b->nb, index, i),
- vtn_vector_insert(b, src, insert, i), dest);
-
- return dest;
-}
-
static nir_ssa_def *
vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
nir_ssa_def *src0, nir_ssa_def *src1,
struct vtn_ssa_value *cur = dest;
unsigned i;
for (i = 0; i < num_indices - 1; i++) {
+ /* If we got a vector here, that means the next index will be trying to
+ * dereference a scalar.
+ */
+ vtn_fail_if(glsl_type_is_vector_or_scalar(cur->type),
+ "OpCompositeInsert has too many indices.");
+ vtn_fail_if(indices[i] >= glsl_get_length(cur->type),
+ "All indices in an OpCompositeInsert must be in-bounds");
cur = cur->elems[indices[i]];
}
if (glsl_type_is_vector_or_scalar(cur->type)) {
+ vtn_fail_if(indices[i] >= glsl_get_vector_elements(cur->type),
+ "All indices in an OpCompositeInsert must be in-bounds");
+
/* According to the SPIR-V spec, OpCompositeInsert may work down to
* the component granularity. In that case, the last index will be
* the index to insert the scalar into the vector.
*/
- cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
+ cur->def = nir_vector_insert_imm(&b->nb, cur->def, insert->def, indices[i]);
} else {
+ vtn_fail_if(indices[i] >= glsl_get_length(cur->type),
+ "All indices in an OpCompositeInsert must be in-bounds");
cur->elems[indices[i]] = insert;
}
for (unsigned i = 0; i < num_indices; i++) {
if (glsl_type_is_vector_or_scalar(cur->type)) {
vtn_assert(i == num_indices - 1);
+ vtn_fail_if(indices[i] >= glsl_get_vector_elements(cur->type),
+ "All indices in an OpCompositeExtract must be in-bounds");
+
/* According to the SPIR-V spec, OpCompositeExtract may work down to
* the component granularity. The last index will be the index of the
* vector to extract.
struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
- ret->def = vtn_vector_extract(b, cur->def, indices[i]);
+ ret->def = nir_channel(&b->nb, cur->def, indices[i]);
return ret;
} else {
+ vtn_fail_if(indices[i] >= glsl_get_length(cur->type),
+ "All indices in an OpCompositeExtract must be in-bounds");
cur = cur->elems[indices[i]];
}
}
switch (opcode) {
case SpvOpVectorExtractDynamic:
- ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
- vtn_ssa_value(b, w[4])->def);
+ ssa->def = nir_vector_extract(&b->nb, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def);
break;
case SpvOpVectorInsertDynamic:
- ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
- vtn_ssa_value(b, w[4])->def,
- vtn_ssa_value(b, w[5])->def);
+ ssa->def = nir_vector_insert(&b->nb, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def,
+ vtn_ssa_value(b, w[5])->def);
break;
case SpvOpVectorShuffle:
break;
case SpvOpCopyLogical:
- case SpvOpCopyObject:
ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
break;
+ case SpvOpCopyObject:
+ vtn_copy_value(b, w[3], w[2]);
+ return;
default:
vtn_fail_with_opcode("unknown composite operation", opcode);
nir_builder_instr_insert(&b->nb, &intrin->instr);
}
-static void
+void
vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope,
SpvMemorySemanticsMask semantics)
{
+ if (b->shader->options->use_scoped_memory_barrier) {
+ vtn_emit_scoped_memory_barrier(b, scope, semantics);
+ return;
+ }
+
static const SpvMemorySemanticsMask all_memory_semantics =
SpvMemorySemanticsUniformMemoryMask |
SpvMemorySemanticsWorkgroupMemoryMask |
SpvMemorySemanticsAtomicCounterMemoryMask |
- SpvMemorySemanticsImageMemoryMask;
+ SpvMemorySemanticsImageMemoryMask |
+ SpvMemorySemanticsOutputMemoryMask;
/* If we're not actually doing a memory barrier, bail */
if (!(semantics & all_memory_semantics))
/* There's only two scopes thing left */
vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice);
- if ((semantics & all_memory_semantics) == all_memory_semantics) {
- vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
- return;
+ /* Map the GLSL memoryBarrier() construct to the corresponding NIR one. */
+ static const SpvMemorySemanticsMask glsl_memory_barrier =
+ SpvMemorySemanticsUniformMemoryMask |
+ SpvMemorySemanticsWorkgroupMemoryMask |
+ SpvMemorySemanticsImageMemoryMask;
+ if ((semantics & glsl_memory_barrier) == glsl_memory_barrier) {
+ vtn_emit_barrier(b, nir_intrinsic_memory_barrier);
+ semantics &= ~(glsl_memory_barrier | SpvMemorySemanticsAtomicCounterMemoryMask);
}
/* Issue a bunch of more specific barriers */
case SpvMemorySemanticsImageMemoryMask:
vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image);
break;
+ case SpvMemorySemanticsOutputMemoryMask:
+ if (b->nb.shader->info.stage == MESA_SHADER_TESS_CTRL)
+ vtn_emit_barrier(b, nir_intrinsic_memory_barrier_tcs_patch);
+ break;
default:
break;;
}
static void
vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
- const uint32_t *w, unsigned count)
+ const uint32_t *w, UNUSED unsigned count)
{
switch (opcode) {
case SpvOpEmitVertex:
}
case SpvOpControlBarrier: {
+ SpvScope execution_scope = vtn_constant_uint(b, w[1]);
SpvScope memory_scope = vtn_constant_uint(b, w[2]);
SpvMemorySemanticsMask memory_semantics = vtn_constant_uint(b, w[3]);
+
+ /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with
+ * memory semantics of None for GLSL barrier().
+ * And before that, prior to c3f1cdfa, emitted the OpControlBarrier with
+ * Device instead of Workgroup for execution scope.
+ */
+ if (b->wa_glslang_cs_barrier &&
+ b->nb.shader->info.stage == MESA_SHADER_COMPUTE &&
+ (execution_scope == SpvScopeWorkgroup ||
+ execution_scope == SpvScopeDevice) &&
+ memory_semantics == SpvMemorySemanticsMaskNone) {
+ execution_scope = SpvScopeWorkgroup;
+ memory_scope = SpvScopeWorkgroup;
+ memory_semantics = SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsWorkgroupMemoryMask;
+ }
+
+ /* From the SPIR-V spec:
+ *
+ * "When used with the TessellationControl execution model, it also
+ * implicitly synchronizes the Output Storage Class: Writes to Output
+ * variables performed by any invocation executed prior to a
+ * OpControlBarrier will be visible to any other invocation after
+ * return from that OpControlBarrier."
+ */
+ if (b->nb.shader->info.stage == MESA_SHADER_TESS_CTRL) {
+ memory_semantics &= ~(SpvMemorySemanticsAcquireMask |
+ SpvMemorySemanticsReleaseMask |
+ SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsSequentiallyConsistentMask);
+ memory_semantics |= SpvMemorySemanticsAcquireReleaseMask |
+ SpvMemorySemanticsOutputMemoryMask;
+ }
+
vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
- SpvScope execution_scope = vtn_constant_uint(b, w[1]);
if (execution_scope == SpvScopeWorkgroup)
- vtn_emit_barrier(b, nir_intrinsic_barrier);
+ vtn_emit_barrier(b, nir_intrinsic_control_barrier);
break;
}
case SpvCapabilityInputAttachment:
case SpvCapabilityImageGatherExtended:
case SpvCapabilityStorageImageExtendedFormats:
+ case SpvCapabilityVector16:
break;
case SpvCapabilityLinkage:
- case SpvCapabilityVector16:
case SpvCapabilityFloat16Buffer:
case SpvCapabilitySparseResidency:
vtn_warn("Unsupported SPIR-V capability: %s",
spv_check_supported(float_controls, cap);
break;
- case SpvCapabilityPhysicalStorageBufferAddressesEXT:
+ case SpvCapabilityPhysicalStorageBufferAddresses:
spv_check_supported(physical_storage_buffer_address, cap);
break;
case SpvCapabilityShaderClockKHR:
spv_check_supported(shader_clock, cap);
+ break;
+
+ case SpvCapabilityVulkanMemoryModel:
+ spv_check_supported(vk_memory_model, cap);
+ break;
+
+ case SpvCapabilityVulkanMemoryModelDeviceScope:
+ spv_check_supported(vk_memory_model_device_scope, cap);
+ break;
+
+ case SpvCapabilityImageReadWriteLodAMD:
+ spv_check_supported(amd_image_read_write_lod, cap);
+ break;
+
+ case SpvCapabilityIntegerFunctions2INTEL:
+ spv_check_supported(integer_functions2, cap);
+ break;
+
+ case SpvCapabilityFragmentMaskAMD:
+ spv_check_supported(amd_fragment_mask, cap);
break;
default:
b->options->temp_addr_format = nir_address_format_64bit_global;
break;
case SpvAddressingModelLogical:
- vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES,
+ vtn_fail_if(b->shader->info.stage == MESA_SHADER_KERNEL,
"AddressingModelLogical only supported for shaders");
- b->shader->info.cs.ptr_size = 0;
b->physical_ptrs = false;
break;
- case SpvAddressingModelPhysicalStorageBuffer64EXT:
+ case SpvAddressingModelPhysicalStorageBuffer64:
vtn_fail_if(!b->options ||
!b->options->caps.physical_storage_buffer_address,
- "AddressingModelPhysicalStorageBuffer64EXT not supported");
+ "AddressingModelPhysicalStorageBuffer64 not supported");
break;
default:
vtn_fail("Unknown addressing model: %s (%u)",
break;
}
- vtn_assert(w[2] == SpvMemoryModelSimple ||
- w[2] == SpvMemoryModelGLSL450 ||
- w[2] == SpvMemoryModelOpenCL);
+ switch (w[2]) {
+ case SpvMemoryModelSimple:
+ case SpvMemoryModelGLSL450:
+ case SpvMemoryModelOpenCL:
+ break;
+ case SpvMemoryModelVulkan:
+ vtn_fail_if(!b->options->caps.vk_memory_model,
+ "Vulkan memory model is unsupported by this driver");
+ break;
+ default:
+ vtn_fail("Unsupported memory model: %s",
+ spirv_memorymodel_to_string(w[2]));
+ break;
+ }
break;
case SpvOpEntryPoint:
vtn_handle_decoration(b, opcode, w, count);
break;
+ case SpvOpExtInst: {
+ struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+ if (val->ext_handler == vtn_handle_non_semantic_instruction) {
+ /* NonSemantic extended instructions are acceptable in preamble. */
+ vtn_handle_non_semantic_instruction(b, w[4], w, count);
+ return true;
+ } else {
+ return false; /* End of preamble. */
+ }
+ }
+
default:
return false; /* End of preamble */
}
static void
vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
- const struct vtn_decoration *mode, void *data)
+ const struct vtn_decoration *mode, UNUSED void *data)
{
vtn_assert(b->entry_point == entry_point);
vtn_handle_variables(b, opcode, w, count);
break;
+ case SpvOpExtInst: {
+ struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+ /* NonSemantic extended instructions are acceptable in preamble, others
+ * will indicate the end of preamble.
+ */
+ return val->ext_handler == vtn_handle_non_semantic_instruction;
+ }
+
default:
return false; /* End of preamble */
}
break;
}
+ case SpvOpFragmentMaskFetchAMD:
+ case SpvOpFragmentFetchAMD:
+ vtn_handle_texture(b, opcode, w, count);
+ break;
+
case SpvOpAtomicLoad:
case SpvOpAtomicExchange:
case SpvOpAtomicCompareExchange:
case SpvOpVectorTimesMatrix:
case SpvOpMatrixTimesVector:
case SpvOpMatrixTimesMatrix:
+ case SpvOpUCountLeadingZerosINTEL:
+ case SpvOpUCountTrailingZerosINTEL:
+ case SpvOpAbsISubINTEL:
+ case SpvOpAbsUSubINTEL:
+ case SpvOpIAddSatINTEL:
+ case SpvOpUAddSatINTEL:
+ case SpvOpIAverageINTEL:
+ case SpvOpUAverageINTEL:
+ case SpvOpIAverageRoundedINTEL:
+ case SpvOpUAverageRoundedINTEL:
+ case SpvOpISubSatINTEL:
+ case SpvOpUSubSatINTEL:
+ case SpvOpIMul32x16INTEL:
+ case SpvOpUMul32x16INTEL:
vtn_handle_alu(b, opcode, w, count);
break;
}
case SpvOpReadClockKHR: {
- assert(vtn_constant_uint(b, w[3]) == SpvScopeSubgroup);
+ SpvScope scope = vtn_constant_uint(b, w[3]);
+ nir_scope nir_scope;
+
+ switch (scope) {
+ case SpvScopeDevice:
+ nir_scope = NIR_SCOPE_DEVICE;
+ break;
+ case SpvScopeSubgroup:
+ nir_scope = NIR_SCOPE_SUBGROUP;
+ break;
+ default:
+ vtn_fail("invalid read clock scope");
+ }
/* Operation supports two result types: uvec2 and uint64_t. The NIR
* intrinsic gives uvec2, so pack the result for the other case.
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_shader_clock);
nir_ssa_dest_init(&intrin->instr, &intrin->dest, 2, 32, NULL);
+ nir_intrinsic_set_memory_scope(intrin, nir_scope);
nir_builder_instr_insert(&b->nb, &intrin->instr);
struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
break;
}
+ case SpvOpLifetimeStart:
+ case SpvOpLifetimeStop:
+ break;
+
default:
vtn_fail_with_opcode("Unhandled opcode", opcode);
}
b->file = NULL;
b->line = -1;
b->col = -1;
- exec_list_make_empty(&b->functions);
+ list_inithead(&b->functions);
b->entry_point_stage = stage;
b->entry_point_name = entry_point_name;
b->options = dup_options;
*/
b->wa_glslang_179 = (generator_id == 8 && generator_version == 1);
+ /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed
+ * to provide correct memory semantics on compute shader barrier()
+ * commands. Prior to that, we need to fix them up ourselves. This
+ * GLSLang fix caused them to bump to generator version 3.
+ */
+ b->wa_glslang_cs_barrier = (generator_id == 8 && generator_version < 3);
+
/* words[2] == generator magic */
unsigned value_id_bound = words[3];
if (words[4] != 0) {
}
/* Set shader info defaults */
- b->shader->info.gs.invocations = 1;
+ if (stage == MESA_SHADER_GEOMETRY)
+ b->shader->info.gs.invocations = 1;
/* Parse rounding mode execution modes. This has to happen earlier than
* other changes in the execution modes since they can affect, for example,
bool progress;
do {
progress = false;
- foreach_list_typed(struct vtn_function, func, node, &b->functions) {
+ vtn_foreach_cf_node(node, &b->functions) {
+ struct vtn_function *func = vtn_cf_node_as_function(node);
if (func->referenced && !func->emitted) {
b->const_table = _mesa_pointer_hash_table_create(b);
* right away. In order to do so, we must lower any constant initializers
* on outputs so nir_remove_dead_variables sees that they're written to.
*/
- nir_lower_constant_initializers(b->shader, nir_var_shader_out);
+ nir_lower_variable_initializers(b->shader, nir_var_shader_out);
nir_remove_dead_variables(b->shader,
nir_var_shader_in | nir_var_shader_out);