struct ac_shader_abi *abi;
gl_shader_stage stage;
+ shader_info *info;
LLVMValueRef *ssa_defs;
enum ac_image_dim dim = get_ac_sampler_dim(ctx, sdim, is_array);
if (dim == ac_image_cube ||
- (ctx->chip_class <= VI && dim == ac_image_3d))
+ (ctx->chip_class <= GFX8 && dim == ac_image_3d))
dim = ac_image_2darray;
return dim;
src0 = ac_to_float(ctx, src0);
result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
- if (ctx->chip_class >= VI) {
+ if (ctx->chip_class >= GFX8) {
LLVMValueRef args[2];
/* Check if the result is a denormal - and flush to 0 if so. */
args[0] = result;
/* need to convert back up to f32 */
result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
- if (ctx->chip_class >= VI)
+ if (ctx->chip_class >= GFX8)
result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
else {
- /* for SI/CIK */
+ /* for GFX6-GFX7 */
/* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
* so compare the result and flush to 0 if it's smaller.
*/
src[i] = get_alu_src(ctx, instr->src[i], src_components);
switch (instr->op) {
- case nir_op_fmov:
- case nir_op_imov:
+ case nir_op_mov:
result = src[0];
break;
case nir_op_fneg:
LLVMBuildExtractElement(ctx->ac.builder, descriptor,
LLVMConstInt(ctx->ac.i32, 2, false), "");
- /* VI only */
- if (ctx->ac.chip_class == VI && in_elements) {
- /* On VI, the descriptor contains the size in bytes,
+ /* GFX8 only */
+ if (ctx->ac.chip_class == GFX8 && in_elements) {
+ /* On GFX8, the descriptor contains the size in bytes,
* but TXQ must return the size in elements.
* The stride is always non-zero for resources using TXQ.
*/
LLVMConstInt(ctx->i32, 0x14000000, false), "");
/* replace the NUM FORMAT in the descriptor */
- tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false), "");
+ tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false), "");
tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
break;
}
- if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= VI) {
+ if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
const struct glsl_type *type = glsl_without_array(var->type);
}
args->attributes = AC_FUNC_ATTR_READNONE;
+ bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
+ ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
+ if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
+ /* Prevent texture instructions with implicit derivatives from being
+ * sinked into branches. */
+ switch (instr->op) {
+ case nir_texop_tex:
+ case nir_texop_txb:
+ case nir_texop_lod:
+ args->attributes |= AC_FUNC_ATTR_CONVERGENT;
+ break;
+ default:
+ break;
+ }
+ }
+
return ac_build_image_opcode(&ctx->ac, args);
}
}
}
- ptr = ac_build_gep0(&ctx->ac, ctx->abi->push_constants, addr);
+ ptr = LLVMBuildGEP(ctx->ac.builder, ctx->abi->push_constants, &addr, 1, "");
if (instr->dest.ssa.bit_size == 8) {
unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
{
unsigned cache_policy = 0;
- /* SI has a TC L1 bug causing corruption of 8bit/16bit stores. All
+ /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
* store opcodes not aligned to a dword are affected. The only way to
* get unaligned stores is through shader images.
*/
- if (((may_store_unaligned && ctx->ac.chip_class == SI) ||
+ if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
/* If this is write-only, don't keep data in L1 to prevent
* evicting L1 cache lines that may be needed by other
* instructions.
u_bit_scan_consecutive_range(&writemask, &start, &count);
- /* Due to an LLVM limitation, split 3-element writes
- * into a 2-element and a 1-element write. */
- if (count == 3) {
+ /* Due to an LLVM limitation with LLVM < 9, split 3-element
+ * writes into a 2-element and a 1-element write. */
+ if (count == 3 &&
+ (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
writemask |= 1 << (start + 2);
count = 2;
}
if (num_bytes == 1) {
ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
offset, ctx->ac.i32_0,
- cache_policy & ac_glc,
- writeonly_memory);
+ cache_policy & ac_glc);
} else if (num_bytes == 2) {
ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
offset, ctx->ac.i32_0,
- cache_policy & ac_glc,
- writeonly_memory);
+ cache_policy & ac_glc);
} else {
int num_channels = num_bytes / 4;
case 16: /* v4f32 */
data_type = ctx->ac.v4f32;
break;
+ case 12: /* v3f32 */
+ data_type = ctx->ac.v3f32;
+ break;
case 8: /* v2f32 */
data_type = ctx->ac.v2f32;
break;
num_channels, offset,
ctx->ac.i32_0, 0,
cache_policy & ac_glc,
- false, writeonly_memory,
- false);
+ false, false);
}
}
}
cache_policy & ac_glc);
} else {
int num_channels = util_next_power_of_two(load_bytes) / 4;
+ bool can_speculate = access & ACCESS_CAN_REORDER;
ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
vindex, offset, immoffset, 0,
cache_policy & ac_glc, 0,
- false, false);
+ can_speculate, false);
}
LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
}
static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
- const nir_intrinsic_instr *instr, bool write)
+ const nir_intrinsic_instr *instr,
+ bool write, bool atomic)
{
LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write);
- if (ctx->abi->gfx9_stride_size_workaround) {
+ if (ctx->abi->gfx9_stride_size_workaround ||
+ (ctx->abi->gfx9_stride_size_workaround_for_atomic && atomic)) {
LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
unsigned num_channels = util_last_bit(mask);
LLVMValueRef rsrc, vindex;
- rsrc = get_image_buffer_descriptor(ctx, instr, false);
+ rsrc = get_image_buffer_descriptor(ctx, instr, false, false);
vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
ctx->ac.i32_0, "");
- /* TODO: set "can_speculate" when OpenGL needs it. */
+ bool can_speculate = access & ACCESS_CAN_REORDER;
res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
ctx->ac.i32_0, num_channels,
!!(args.cache_policy & ac_glc),
- false);
+ can_speculate);
res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
args.cache_policy = get_cache_policy(ctx, access, true, writeonly_memory);
if (dim == GLSL_SAMPLER_DIM_BUF) {
- LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true);
+ LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true, false);
LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
unsigned src_channels = ac_get_llvm_num_components(src);
LLVMValueRef vindex;
ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
ctx->ac.i32_0, src_channels,
- args.cache_policy & ac_glc,
- writeonly_memory);
+ args.cache_policy & ac_glc, false);
} else {
args.opcode = ac_image_store;
args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
params[param_count++] = get_src(ctx, instr->src[3]);
if (dim == GLSL_SAMPLER_DIM_BUF) {
- params[param_count++] = get_image_buffer_descriptor(ctx, instr, true);
+ params[param_count++] = get_image_buffer_descriptor(ctx, instr, true, true);
params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
ctx->ac.i32_0, ""); /* vindex */
params[param_count++] = ctx->ac.i32_0; /* voffset */
void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
{
- /* SI only (thanks to a hw bug workaround):
+ /* GFX6 only (thanks to a hw bug workaround):
* The real barrier instruction isn’t needed, because an entire patch
* always fits into a single wave.
*/
- if (ac->chip_class == SI && stage == MESA_SHADER_TESS_CTRL) {
+ if (ac->chip_class == GFX6 && stage == MESA_SHADER_TESS_CTRL) {
ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
return;
}
case nir_intrinsic_quad_swap_diagonal:
result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
break;
+ case nir_intrinsic_quad_swizzle_amd: {
+ uint32_t mask = nir_intrinsic_swizzle_mask(instr);
+ result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
+ mask & 0x3, (mask >> 2) & 0x3,
+ (mask >> 4) & 0x3, (mask >> 6) & 0x3);
+ break;
+ }
+ case nir_intrinsic_masked_swizzle_amd: {
+ uint32_t mask = nir_intrinsic_swizzle_mask(instr);
+ result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
+ break;
+ }
+ case nir_intrinsic_write_invocation_amd:
+ result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
+ get_src(ctx, instr->src[1]),
+ get_src(ctx, instr->src[2]));
+ break;
+ case nir_intrinsic_mbcnt_amd:
+ result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
+ break;
default:
fprintf(stderr, "Unknown intrinsic: ");
nir_print_instr(&instr->instr, stderr);
/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
*
- * SI-CI:
+ * GFX6-GFX7:
* If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
* filtering manually. The driver sets img7 to a mask clearing
* MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
* s_and_b32 samp0, samp0, img7
*
- * VI:
+ * GFX8:
* The ANISO_OVERRIDE sampler field enables this fix in TA.
*/
static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef img7, samp0;
- if (ctx->ac.chip_class >= VI)
+ if (ctx->ac.chip_class >= GFX8)
return samp;
img7 = LLVMBuildExtractElement(builder, res,
* It's unnecessary if the original texture format was
* Z32_FLOAT, but we don't know that here.
*/
- if (args.compare && ctx->ac.chip_class >= VI && ctx->abi->clamp_shadow_reference)
+ if (args.compare && ctx->ac.chip_class >= GFX8 && ctx->abi->clamp_shadow_reference)
args.compare = ac_build_clamp(&ctx->ac, ac_to_float(&ctx->ac, args.compare));
/* pack derivatives */
args.offset = NULL;
}
- /* TODO TG4 support */
+ /* DMASK was repurposed for GATHER4. 4 components are always
+ * returned and DMASK works like a swizzle - it selects
+ * the component to fetch. The only valid DMASK values are
+ * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
+ * (red,red,red,red) etc.) The ISA document doesn't mention
+ * this.
+ */
args.dmask = 0xf;
if (instr->op == nir_texop_tg4) {
if (instr->is_shadow)
ctx.abi = abi;
ctx.stage = nir->info.stage;
+ ctx.info = &nir->info;
ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
* by the reality that LLVM 5.0 doesn't have working VGPR indexing
* on GFX9.
*/
- bool llvm_has_working_vgpr_indexing = chip_class <= VI;
+ bool llvm_has_working_vgpr_indexing = chip_class <= GFX8;
/* TODO: Indirect indexing of GS inputs is unimplemented.
*
projects / mesa.git / blobdiff