[Mesa-dev] [PATCH v6 31/34] nvir/nir: implement images
Karol Herbst
kherbst at redhat.com
Mon Mar 19 15:44:24 UTC 2018
v3: fix compiler warnings
v4: use loadFrom helper
v5: fix signed min/max
v6: set tex mask
add support for indirect image access
set cache mode
Signed-off-by: Karol Herbst <kherbst at redhat.com>
---
.../drivers/nouveau/codegen/nv50_ir_from_nir.cpp | 395 +++++++++++++++++++--
1 file changed, 375 insertions(+), 20 deletions(-)
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
index e15f1734cc3..5c3fde32601 100644
--- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
+++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
@@ -91,6 +91,8 @@ private:
LValues& convert(nir_register *);
LValues& convert(nir_ssa_def *);
+ ImgFormat convertGLImgFormat(GLuint);
+
Value* getSrc(nir_alu_src *, uint8_t component = 0);
Value* getSrc(nir_register *, uint8_t);
Value* getSrc(nir_src *, uint8_t, bool indirect = false);
@@ -120,6 +122,7 @@ private:
DataType getDType(nir_alu_instr*);
DataType getDType(nir_intrinsic_instr*);
+ DataType getDType(nir_intrinsic_instr*, bool isSigned);
DataType getDType(nir_op, NirSSADefBitSize);
std::vector<DataType> getSTypes(nir_alu_instr*);
@@ -153,6 +156,11 @@ private:
/* tex stuff */
Value* applyProjection(Value *src, Value *proj);
+ unsigned int getNIRArgCount(TexInstruction::Target&);
+
+ /* image stuff */
+ uint16_t derefImageVar(nir_deref_var *, Value **indirect);
+ CacheMode getCacheModeFromVar(nir_variable *);
nir_shader *nir;
@@ -244,11 +252,30 @@ Converter::getDType(nir_alu_instr *insn)
DataType
Converter::getDType(nir_intrinsic_instr *insn)
+{
+ bool isSigned;
+ switch (insn->intrinsic) {
+ case nir_intrinsic_shared_atomic_imax:
+ case nir_intrinsic_shared_atomic_imin:
+ case nir_intrinsic_ssbo_atomic_imax:
+ case nir_intrinsic_ssbo_atomic_imin:
+ isSigned = true;
+ break;
+ default:
+ isSigned = false;
+ break;
+ }
+
+ return getDType(insn, isSigned);
+}
+
+DataType
+Converter::getDType(nir_intrinsic_instr *insn, bool isSigned)
{
if (insn->dest.is_ssa)
- return typeOfSize(insn->dest.ssa.bit_size / 8, false, false);
+ return typeOfSize(insn->dest.ssa.bit_size / 8, false, isSigned);
else
- return typeOfSize(insn->dest.reg.reg->bit_size / 8, false, false);
+ return typeOfSize(insn->dest.reg.reg->bit_size / 8, false, isSigned);
}
DataType
@@ -445,28 +472,31 @@ Converter::getSubOp(nir_op op)
}
}
+#define CASE_OP_INTR_ATOM(nir, nvir) \
+ case nir_intrinsic_image_atomic_ ## nir : \
+ case nir_intrinsic_shared_atomic_ ## nir : \
+ case nir_intrinsic_ssbo_atomic_ ## nir : \
+ return NV50_IR_SUBOP_ATOM_ ## nvir
+#define CASE_OP_INTR_ATOM_S(nir, nvir) \
+ case nir_intrinsic_shared_atomic_ ## nir : \
+ case nir_intrinsic_ssbo_atomic_ ## nir : \
+ return NV50_IR_SUBOP_ATOM_ ## nvir
int
Converter::getSubOp(nir_intrinsic_op op)
{
switch (op) {
- case nir_intrinsic_ssbo_atomic_add:
- return NV50_IR_SUBOP_ATOM_ADD;
- case nir_intrinsic_ssbo_atomic_and:
- return NV50_IR_SUBOP_ATOM_AND;
- case nir_intrinsic_ssbo_atomic_comp_swap:
- return NV50_IR_SUBOP_ATOM_CAS;
- case nir_intrinsic_ssbo_atomic_exchange:
- return NV50_IR_SUBOP_ATOM_EXCH;
- case nir_intrinsic_ssbo_atomic_or:
- return NV50_IR_SUBOP_ATOM_OR;
- case nir_intrinsic_ssbo_atomic_imax:
- case nir_intrinsic_ssbo_atomic_umax:
- return NV50_IR_SUBOP_ATOM_MAX;
- case nir_intrinsic_ssbo_atomic_imin:
- case nir_intrinsic_ssbo_atomic_umin:
- return NV50_IR_SUBOP_ATOM_MIN;
- case nir_intrinsic_ssbo_atomic_xor:
- return NV50_IR_SUBOP_ATOM_XOR;
+ CASE_OP_INTR_ATOM(add, ADD);
+ CASE_OP_INTR_ATOM(and, AND);
+ CASE_OP_INTR_ATOM(comp_swap, CAS);
+ CASE_OP_INTR_ATOM(exchange, EXCH);
+ CASE_OP_INTR_ATOM(or, OR);
+ case nir_intrinsic_image_atomic_max:
+ CASE_OP_INTR_ATOM_S(imax, MAX);
+ CASE_OP_INTR_ATOM_S(umax, MAX);
+ case nir_intrinsic_image_atomic_min:
+ CASE_OP_INTR_ATOM_S(imin, MIN);
+ CASE_OP_INTR_ATOM_S(umin, MIN);
+ CASE_OP_INTR_ATOM(xor, XOR);
case nir_intrinsic_vote_all:
return NV50_IR_SUBOP_VOTE_ALL;
case nir_intrinsic_vote_any:
@@ -479,6 +509,8 @@ Converter::getSubOp(nir_intrinsic_op op)
return 0;
}
}
+#undef CASE_OP_INTR_ATOM
+#undef CASE_OP_INTR_ATOM_S
CondCode
Converter::getCondCode(nir_op op)
@@ -1628,6 +1660,68 @@ Converter::convert(nir_intrinsic_op intr)
}
}
+ImgFormat
+Converter::convertGLImgFormat(GLuint format)
+{
+#define FMT_CASE(a, b) \
+ case GL_ ## a: return nv50_ir::FMT_ ## b
+
+ switch (format) {
+ FMT_CASE(NONE, NONE);
+
+ FMT_CASE(RGBA32F, RGBA32F);
+ FMT_CASE(RGBA16F, RGBA16F);
+ FMT_CASE(RG32F, RG32F);
+ FMT_CASE(RG16F, RG16F);
+ FMT_CASE(R11F_G11F_B10F, R11G11B10F);
+ FMT_CASE(R32F, R32F);
+ FMT_CASE(R16F, R16F);
+
+ FMT_CASE(RGBA32UI, RGBA32UI);
+ FMT_CASE(RGBA16UI, RGBA16UI);
+ FMT_CASE(RGB10_A2UI, RGB10A2UI);
+ FMT_CASE(RGBA8UI, RGBA8UI);
+ FMT_CASE(RG32UI, RG32UI);
+ FMT_CASE(RG16UI, RG16UI);
+ FMT_CASE(RG8UI, RG8UI);
+ FMT_CASE(R32UI, R32UI);
+ FMT_CASE(R16UI, R16UI);
+ FMT_CASE(R8UI, R8UI);
+
+ FMT_CASE(RGBA32I, RGBA32I);
+ FMT_CASE(RGBA16I, RGBA16I);
+ FMT_CASE(RGBA8I, RGBA8I);
+ FMT_CASE(RG32I, RG32I);
+ FMT_CASE(RG16I, RG16I);
+ FMT_CASE(RG8I, RG8I);
+ FMT_CASE(R32I, R32I);
+ FMT_CASE(R16I, R16I);
+ FMT_CASE(R8I, R8I);
+
+ FMT_CASE(RGBA16, RGBA16);
+ FMT_CASE(RGB10_A2, RGB10A2);
+ FMT_CASE(RGBA8, RGBA8);
+ FMT_CASE(RG16, RG16);
+ FMT_CASE(RG8, RG8);
+ FMT_CASE(R16, R16);
+ FMT_CASE(R8, R8);
+
+ FMT_CASE(RGBA16_SNORM, RGBA16_SNORM);
+ FMT_CASE(RGBA8_SNORM, RGBA8_SNORM);
+ FMT_CASE(RG16_SNORM, RG16_SNORM);
+ FMT_CASE(RG8_SNORM, RG8_SNORM);
+ FMT_CASE(R16_SNORM, R16_SNORM);
+ FMT_CASE(R8_SNORM, R8_SNORM);
+
+ FMT_CASE(BGRA_INTEGER, BGRA8);
+ default:
+ ERROR("unknown format %x\n", format);
+ assert(false);
+ return nv50_ir::FMT_NONE;
+ }
+#undef FMT_CASE
+}
+
bool
Converter::visit(nir_intrinsic_instr *insn)
{
@@ -1966,6 +2060,28 @@ Converter::visit(nir_intrinsic_instr *insn)
info->io.globalAccess |= 0x1;
break;
}
+ case nir_intrinsic_shared_atomic_add:
+ case nir_intrinsic_shared_atomic_and:
+ case nir_intrinsic_shared_atomic_comp_swap:
+ case nir_intrinsic_shared_atomic_exchange:
+ case nir_intrinsic_shared_atomic_or:
+ case nir_intrinsic_shared_atomic_imax:
+ case nir_intrinsic_shared_atomic_imin:
+ case nir_intrinsic_shared_atomic_umax:
+ case nir_intrinsic_shared_atomic_umin:
+ case nir_intrinsic_shared_atomic_xor: {
+ const DataType dType = getDType(insn);
+ LValues &newDefs = convert(&insn->dest);
+ Value *indirectOffset;
+ uint32_t offset = getIndirect(&insn->src[0], 0, &indirectOffset);
+ Symbol *sym = mkSymbol(FILE_MEMORY_SHARED, 0, dType, offset);
+ Instruction *atom = mkOp2(OP_ATOM, dType, newDefs[0], sym, getSrc(&insn->src[1], 0));
+ if (op == nir_intrinsic_shared_atomic_comp_swap)
+ atom->setSrc(2, getSrc(&insn->src[2], 0));
+ atom->setIndirect(0, 0, indirectOffset);
+ atom->subOp = getSubOp(op);
+ break;
+ }
case nir_intrinsic_ssbo_atomic_add:
case nir_intrinsic_ssbo_atomic_and:
case nir_intrinsic_ssbo_atomic_comp_swap:
@@ -1995,6 +2111,182 @@ Converter::visit(nir_intrinsic_instr *insn)
info->io.globalAccess |= 0x2;
break;
}
+ case nir_intrinsic_image_store: {
+ nir_variable *tex = insn->variables[0]->var;
+ const glsl_type *type = tex->type->without_array();
+ TexInstruction::Target target =
+ convert((glsl_sampler_dim)type->sampler_dimensionality,
+ type->sampler_array, type->sampler_shadow);
+ std::vector<Value*> srcs, dummy;
+ unsigned int argCount = getNIRArgCount(target);
+ auto location = tex->data.driver_location;
+ Value *indirect = nullptr;
+
+ srcs.resize(argCount + 4);
+ for (auto i = 0u; i < argCount; ++i)
+ srcs[i] = getSrc(&insn->src[0], i);
+ for (auto i = 0u; i < 4; ++i)
+ srcs[argCount + i] = getSrc(&insn->src[2], i);
+
+ if (location == 0xffffffff)
+ location = derefImageVar(insn->variables[0], &indirect);
+
+ TexInstruction *texi = mkTex(OP_SUSTP, target.getEnum(), location, 0, dummy, srcs);
+ texi->tex.format =
+ &nv50_ir::TexInstruction::formatTable[convertGLImgFormat(tex->data.image.format)];
+ texi->tex.mask = 0xf;
+ texi->cache = getCacheModeFromVar(tex);
+
+ if (indirect)
+ texi->setIndirectR(indirect);
+
+ info->io.globalAccess |= 0x2;
+ break;
+ }
+ case nir_intrinsic_image_load: {
+ nir_variable *tex = insn->variables[0]->var;
+ const glsl_type *type = tex->type->without_array();
+ TexInstruction::Target target =
+ convert((glsl_sampler_dim)type->sampler_dimensionality,
+ type->sampler_array, type->sampler_shadow);
+ std::vector<Value*> srcs, defs;
+ uint32_t mask = 0;
+ LValues &newDefs = convert(&insn->dest);
+ unsigned int argCount = getNIRArgCount(target);
+ auto location = tex->data.driver_location;
+ Value *indirect = nullptr;
+
+ defs.resize(newDefs.size());
+ for (auto i = 0u; i < newDefs.size(); ++i) {
+ defs[i] = newDefs[i];
+ mask |= 1 << i;
+ }
+
+ srcs.resize(argCount);
+ for (auto i = 0u; i < argCount; ++i)
+ srcs[i] = getSrc(&insn->src[0], i);
+
+ if (location == 0xffffffff)
+ location = derefImageVar(insn->variables[0], &indirect);
+
+ TexInstruction *texi = mkTex(OP_SULDP, target.getEnum(), location, 0, defs, srcs);
+ texi->tex.format =
+ &nv50_ir::TexInstruction::formatTable[convertGLImgFormat(tex->data.image.format)];
+ texi->tex.mask = mask;
+ texi->cache = getCacheModeFromVar(tex);
+
+ if (indirect)
+ texi->setIndirectR(indirect);
+
+ info->io.globalAccess |= 0x1;
+ break;
+ }
+ case nir_intrinsic_image_samples:
+ case nir_intrinsic_image_size: {
+ nir_variable *tex = insn->variables[0]->var;
+ const glsl_type *type = tex->type->without_array();
+ TexInstruction::Target target =
+ convert((glsl_sampler_dim)type->sampler_dimensionality,
+ type->sampler_array, type->sampler_shadow);
+ std::vector<Value*> srcs, defs;
+ uint32_t mask = 0;
+ LValues &newDefs = convert(&insn->dest);
+ auto location = tex->data.driver_location;
+ Value *indirect = nullptr;
+
+ defs.resize(newDefs.size());
+ for (auto i = 0u; i < newDefs.size(); ++i) {
+ defs[i] = newDefs[i];
+ mask |= 1 << i;
+ }
+
+ if (location == 0xffffffff)
+ location = derefImageVar(insn->variables[0], &indirect);
+
+ TexInstruction *texi = mkTex(OP_SUQ, target.getEnum(), location, 0, defs, srcs);
+ texi->tex.mask = mask;
+
+ if (indirect)
+ texi->setIndirectR(indirect);
+
+ break;
+ }
+ case nir_intrinsic_image_atomic_add:
+ case nir_intrinsic_image_atomic_and:
+ case nir_intrinsic_image_atomic_comp_swap:
+ case nir_intrinsic_image_atomic_exchange:
+ case nir_intrinsic_image_atomic_max:
+ case nir_intrinsic_image_atomic_min:
+ case nir_intrinsic_image_atomic_or:
+ case nir_intrinsic_image_atomic_xor: {
+ nir_variable *tex = insn->variables[0]->var;
+ const glsl_type *type = tex->type->without_array();
+ const DataType dType = getDType(insn, type->base_type == GLSL_TYPE_INT);
+ TexInstruction::Target target =
+ convert((glsl_sampler_dim)type->sampler_dimensionality,
+ type->sampler_array, type->sampler_shadow);
+ std::vector<Value*> srcs, defs;
+ LValues &newDefs = convert(&insn->dest);
+ unsigned int argCount = getNIRArgCount(target);
+ auto location = tex->data.driver_location;
+ Value *indirect = nullptr;
+
+ defs.push_back(newDefs[0]);
+
+ srcs.resize(argCount);
+ for (auto i = 0u; i < argCount; ++i)
+ srcs[i] = getSrc(&insn->src[0], i);
+ srcs.push_back(getSrc(&insn->src[2], 0));
+ if (op == nir_intrinsic_image_atomic_comp_swap)
+ srcs.push_back(getSrc(&insn->src[3], 0));
+
+ if (location == 0xffffffff)
+ location = derefImageVar(insn->variables[0], &indirect);
+
+ TexInstruction *texi =
+ mkTex(OP_SUREDP, target.getEnum(), location, 0, defs, srcs);
+ texi->tex.mask = 1;
+ texi->setType(dType);
+ texi->subOp = getSubOp(op);
+
+ if (indirect)
+ texi->setIndirectR(indirect);
+
+ info->io.globalAccess |= 0x2;
+ break;
+ }
+ case nir_intrinsic_store_shared: {
+ DataType sType = getSType(insn->src[0], false, false);
+ Value *indirectOffset;
+ uint32_t offset = getIndirect(&insn->src[1], 0, &indirectOffset);
+
+ for (auto i = 0u; i < insn->num_components; ++i) {
+ if (!((1u << i) & nir_intrinsic_write_mask(insn)))
+ continue;
+ Symbol *sym = mkSymbol(FILE_MEMORY_SHARED, 0, sType, offset + i * typeSizeof(sType));
+ mkStore(OP_STORE, sType, sym, indirectOffset, getSrc(&insn->src[0], i));
+ }
+ break;
+ }
+ case nir_intrinsic_load_shared: {
+ const DataType dType = getDType(insn);
+ LValues &newDefs = convert(&insn->dest);
+ Value *indirectOffset;
+ uint32_t offset = getIndirect(&insn->src[0], 0, &indirectOffset);
+
+ for (auto i = 0u; i < insn->num_components; ++i)
+ loadFrom(FILE_MEMORY_SHARED, 0, dType, newDefs[i], offset, i, indirectOffset);
+
+ break;
+ }
+ case nir_intrinsic_barrier: {
+ // TODO: add flag to shader_info
+ info->numBarriers = 1;
+ Instruction *bar = mkOp2(OP_BAR, TYPE_U32, NULL, mkImm(0), mkImm(0));
+ bar->fixed = 1;
+ bar->subOp = NV50_IR_SUBOP_BAR_SYNC;
+ break;
+ }
default:
ERROR("unknown nir_intrinsic_op %s\n", nir_intrinsic_infos[op].name);
return false;
@@ -2465,6 +2757,69 @@ Converter::applyProjection(Value *src, Value *proj)
return mkOp2v(OP_MUL, TYPE_F32, getScratch(), src, proj);
}
+unsigned int
+Converter::getNIRArgCount(TexInstruction::Target& target)
+{
+ unsigned int result = target.getArgCount();
+ if (target.isCube() && target.isArray())
+ return result - 1;
+ return result;
+}
+
+uint16_t
+Converter::derefImageVar(nir_deref_var *deref, Value **indirect)
+{
+ uint16_t idx = 0;
+
+ if (indirect)
+ *indirect = loadImm(getSSA(), 0u);
+
+ for (nir_deref *child = deref->deref.child; child; child = child->child) {
+ switch (child->deref_type) {
+ case nir_deref_type_array: {
+ nir_deref_array *arr = nir_deref_as_array(child);
+ auto size = type_size(child->type);
+ idx += size * arr->base_offset;
+
+ switch (arr->deref_array_type) {
+ case nir_deref_array_type_direct:
+ break;
+ case nir_deref_array_type_indirect: {
+ Value *offset =
+ mkOp2v(OP_MUL, TYPE_U32, getSSA(), loadImm(getSSA(), size),
+ getSrc(&arr->indirect, 0, false));
+ *indirect = mkOp2v(OP_ADD, TYPE_U32, getSSA(), *indirect, offset);
+ break;
+ }
+ case nir_deref_array_type_wildcard:
+ ERROR("Wildcard derefs not supported!!\n");
+ assert(false);
+ break;
+ default:
+ ERROR("Unknown deref type!\n");
+ assert(false);
+ break;
+ }
+ break;
+ }
+ default:
+ ERROR("Only array derefs supported yet!\n");
+ assert(false);
+ }
+ }
+ return idx;
+}
+
+CacheMode
+Converter::getCacheModeFromVar(nir_variable *var)
+{
+ if (var->data.image._volatile)
+ return CACHE_CV;
+ if (var->data.image.coherent)
+ return CACHE_CG;
+ return CACHE_CA;
+}
+
bool
Converter::visit(nir_tex_instr *insn)
{
--
2.14.3
More information about the mesa-dev
mailing list