[Mesa-dev] [PATCH 02/18] radeonsi: add a workaround for clamping unaligned RGB 8 & 16-bit vertex loads
Marek Olšák
maraeo at gmail.com
Thu Feb 16 12:52:54 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeonsi/si_shader.c | 38 +++++++++++++++++++---
src/gallium/drivers/radeonsi/si_shader.h | 4 +++
src/gallium/drivers/radeonsi/si_shader_internal.h | 3 ++
.../drivers/radeonsi/si_shader_tgsi_setup.c | 3 ++
src/gallium/drivers/radeonsi/si_state.c | 18 ++++++++--
5 files changed, 60 insertions(+), 6 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 8b9fed9..1829e3e 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -361,28 +361,41 @@ static void declare_input_vs(
t_offset = lp_build_const_int32(gallivm, input_index);
t_list = ac_build_indexed_load_const(&ctx->ac, t_list_ptr, t_offset);
vertex_index = LLVMGetParam(ctx->main_fn,
ctx->param_vertex_index0 +
input_index);
fix_fetch = ctx->shader->key.mono.vs.fix_fetch[input_index];
- /* Do multiple loads for double formats. */
- if (fix_fetch == SI_FIX_FETCH_RGB_64_FLOAT) {
+ /* Do multiple loads for special formats. */
+ switch (fix_fetch) {
+ case SI_FIX_FETCH_RGB_64_FLOAT:
num_fetches = 3; /* 3 2-dword loads */
fetch_stride = 8;
- } else if (fix_fetch == SI_FIX_FETCH_RGBA_64_FLOAT) {
+ break;
+ case SI_FIX_FETCH_RGBA_64_FLOAT:
num_fetches = 2; /* 2 4-dword loads */
fetch_stride = 16;
- } else {
+ break;
+ case SI_FIX_FETCH_RGB_8:
+ case SI_FIX_FETCH_RGB_8_INT:
+ num_fetches = 3;
+ fetch_stride = 1;
+ break;
+ case SI_FIX_FETCH_RGB_16:
+ case SI_FIX_FETCH_RGB_16_INT:
+ num_fetches = 3;
+ fetch_stride = 2;
+ break;
+ default:
num_fetches = 1;
fetch_stride = 0;
}
args[0] = t_list;
args[2] = vertex_index;
for (unsigned i = 0; i < num_fetches; i++) {
args[1] = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
@@ -505,20 +518,37 @@ static void declare_input_vs(
out[chan] = extract_double_to_float(ctx, input[chan], 0);
out[3] = LLVMConstReal(ctx->f32, 1);
break;
case SI_FIX_FETCH_RGBA_64_FLOAT:
for (chan = 0; chan < 4; chan++) {
out[chan] = extract_double_to_float(ctx, input[chan / 2],
chan % 2);
}
break;
+ case SI_FIX_FETCH_RGB_8:
+ case SI_FIX_FETCH_RGB_8_INT:
+ case SI_FIX_FETCH_RGB_16:
+ case SI_FIX_FETCH_RGB_16_INT:
+ for (chan = 0; chan < 3; chan++) {
+ out[chan] = LLVMBuildExtractElement(gallivm->builder,
+ input[chan],
+ ctx->i32_0, "");
+ }
+ if (fix_fetch == SI_FIX_FETCH_RGB_8 ||
+ fix_fetch == SI_FIX_FETCH_RGB_16) {
+ out[3] = LLVMConstReal(ctx->f32, 1);
+ } else {
+ out[3] = LLVMBuildBitCast(gallivm->builder, ctx->i32_1,
+ ctx->f32, "");
+ }
+ break;
}
}
static LLVMValueRef get_primitive_id(struct lp_build_tgsi_context *bld_base,
unsigned swizzle)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
if (swizzle > 0)
return bld_base->uint_bld.zero;
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 4616190..da88df0 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -243,20 +243,24 @@ enum {
SI_FIX_FETCH_RGBX_32_UNORM,
SI_FIX_FETCH_RGBA_32_SNORM,
SI_FIX_FETCH_RGBX_32_SNORM,
SI_FIX_FETCH_RGBA_32_USCALED,
SI_FIX_FETCH_RGBA_32_SSCALED,
SI_FIX_FETCH_RGBA_32_FIXED,
SI_FIX_FETCH_RGBX_32_FIXED,
SI_FIX_FETCH_RG_64_FLOAT,
SI_FIX_FETCH_RGB_64_FLOAT,
SI_FIX_FETCH_RGBA_64_FLOAT,
+ SI_FIX_FETCH_RGB_8, /* A = 1.0 */
+ SI_FIX_FETCH_RGB_8_INT, /* A = 1 */
+ SI_FIX_FETCH_RGB_16,
+ SI_FIX_FETCH_RGB_16_INT,
};
struct si_shader;
/* State of the context creating the shader object. */
struct si_compiler_ctx_state {
/* Should only be used by si_init_shader_selector_async and
* si_build_shader_variant if thread_index == -1 (non-threaded). */
LLVMTargetMachineRef tm;
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 26cc28d..8fde6c2 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -142,20 +142,23 @@ struct si_shader_context {
LLVMTypeRef i32;
LLVMTypeRef i64;
LLVMTypeRef i128;
LLVMTypeRef f32;
LLVMTypeRef v16i8;
LLVMTypeRef v2i32;
LLVMTypeRef v4i32;
LLVMTypeRef v4f32;
LLVMTypeRef v8i32;
+ LLVMValueRef i32_0;
+ LLVMValueRef i32_1;
+
LLVMValueRef shared_memory;
};
static inline struct si_shader_context *
si_shader_context(struct lp_build_tgsi_context *bld_base)
{
return (struct si_shader_context*)bld_base;
}
void si_llvm_add_attribute(LLVMValueRef F, const char *name, int value);
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
index c7445e0..c7019c1 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
@@ -1359,20 +1359,23 @@ void si_llvm_context_init(struct si_shader_context *ctx,
ctx->i8 = LLVMInt8TypeInContext(ctx->gallivm.context);
ctx->i32 = LLVMInt32TypeInContext(ctx->gallivm.context);
ctx->i64 = LLVMInt64TypeInContext(ctx->gallivm.context);
ctx->i128 = LLVMIntTypeInContext(ctx->gallivm.context, 128);
ctx->f32 = LLVMFloatTypeInContext(ctx->gallivm.context);
ctx->v16i8 = LLVMVectorType(ctx->i8, 16);
ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
+
+ ctx->i32_0 = LLVMConstInt(ctx->i32, 0, 0);
+ ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
}
void si_llvm_create_func(struct si_shader_context *ctx,
const char *name,
LLVMTypeRef *return_types, unsigned num_return_elems,
LLVMTypeRef *ParamTypes, unsigned ParamCount)
{
LLVMTypeRef main_fn_type, ret_type;
LLVMBasicBlockRef main_fn_body;
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index d9b9f83..024de8b 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -1725,35 +1725,35 @@ static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
/* See whether the components are of the same size. */
for (i = 0; i < desc->nr_channels; i++) {
if (desc->channel[first_non_void].size != desc->channel[i].size)
return V_008F0C_BUF_DATA_FORMAT_INVALID;
}
switch (desc->channel[first_non_void].size) {
case 8:
switch (desc->nr_channels) {
case 1:
+ case 3: /* 3 loads */
return V_008F0C_BUF_DATA_FORMAT_8;
case 2:
return V_008F0C_BUF_DATA_FORMAT_8_8;
- case 3:
case 4:
return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
}
break;
case 16:
switch (desc->nr_channels) {
case 1:
+ case 3: /* 3 loads */
return V_008F0C_BUF_DATA_FORMAT_16;
case 2:
return V_008F0C_BUF_DATA_FORMAT_16_16;
- case 3:
case 4:
return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
}
break;
case 32:
switch (desc->nr_channels) {
case 1:
return V_008F0C_BUF_DATA_FORMAT_32;
case 2:
return V_008F0C_BUF_DATA_FORMAT_32_32;
@@ -3452,20 +3452,34 @@ static void *si_create_vertex_elements(struct pipe_context *ctx,
case 4:
v->fix_fetch[i] = SI_FIX_FETCH_RGBA_64_FLOAT;
swizzle[0] = PIPE_SWIZZLE_X; /* 2 loads */
swizzle[1] = PIPE_SWIZZLE_Y;
swizzle[2] = PIPE_SWIZZLE_Z;
swizzle[3] = PIPE_SWIZZLE_W;
break;
default:
assert(0);
}
+ } else if (channel && desc->nr_channels == 3) {
+ assert(desc->swizzle[0] == PIPE_SWIZZLE_X);
+
+ if (channel->size == 8) {
+ if (channel->pure_integer)
+ v->fix_fetch[i] = SI_FIX_FETCH_RGB_8_INT;
+ else
+ v->fix_fetch[i] = SI_FIX_FETCH_RGB_8;
+ } else if (channel->size == 16) {
+ if (channel->pure_integer)
+ v->fix_fetch[i] = SI_FIX_FETCH_RGB_16_INT;
+ else
+ v->fix_fetch[i] = SI_FIX_FETCH_RGB_16;
+ }
}
v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
S_008F0C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
S_008F0C_NUM_FORMAT(num_format) |
S_008F0C_DATA_FORMAT(data_format);
/* We work around the fact that 8_8_8 and 16_16_16 data formats
--
2.7.4
More information about the mesa-dev
mailing list