* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
-#include "util/u_memory.h"
-#include "tgsi/tgsi_strings.h"
-#include "tgsi/tgsi_from_mesa.h"
-
#include "ac_exp_param.h"
#include "ac_rtld.h"
-#include "si_shader_internal.h"
-#include "si_pipe.h"
-#include "sid.h"
-
#include "compiler/nir/nir.h"
#include "compiler/nir/nir_serialize.h"
+#include "si_pipe.h"
+#include "si_shader_internal.h"
+#include "sid.h"
+#include "tgsi/tgsi_from_mesa.h"
+#include "tgsi/tgsi_strings.h"
+#include "util/u_memory.h"
-static const char scratch_rsrc_dword0_symbol[] =
- "SCRATCH_RSRC_DWORD0";
+static const char scratch_rsrc_dword0_symbol[] = "SCRATCH_RSRC_DWORD0";
-static const char scratch_rsrc_dword1_symbol[] =
- "SCRATCH_RSRC_DWORD1";
+static const char scratch_rsrc_dword1_symbol[] = "SCRATCH_RSRC_DWORD1";
static void si_dump_shader_key(const struct si_shader *shader, FILE *f);
-static void si_build_vs_prolog_function(struct si_shader_context *ctx,
- union si_shader_part_key *key);
-
/** Whether the shader runs as a combination of multiple API shaders */
-static bool is_multi_part_shader(struct si_shader_context *ctx)
+bool si_is_multi_part_shader(struct si_shader *shader)
{
- if (ctx->screen->info.chip_class <= GFX8)
- return false;
+ if (shader->selector->screen->info.chip_class <= GFX8)
+ return false;
- return ctx->shader->key.as_ls ||
- ctx->shader->key.as_es ||
- ctx->type == PIPE_SHADER_TESS_CTRL ||
- ctx->type == PIPE_SHADER_GEOMETRY;
+ return shader->key.as_ls || shader->key.as_es ||
+ shader->selector->info.stage == MESA_SHADER_TESS_CTRL ||
+ shader->selector->info.stage == MESA_SHADER_GEOMETRY;
}
/** Whether the shader runs on a merged HW stage (LSHS or ESGS) */
-bool si_is_merged_shader(struct si_shader_context *ctx)
+bool si_is_merged_shader(struct si_shader *shader)
{
- return ctx->shader->key.as_ngg || is_multi_part_shader(ctx);
+ return shader->key.as_ngg || si_is_multi_part_shader(shader);
}
/**
*/
unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned index)
{
- switch (semantic_name) {
- case TGSI_SEMANTIC_TESSOUTER:
- return 0;
- case TGSI_SEMANTIC_TESSINNER:
- return 1;
- case TGSI_SEMANTIC_PATCH:
- assert(index < 30);
- return 2 + index;
-
- default:
- assert(!"invalid semantic name");
- return 0;
- }
+ switch (semantic_name) {
+ case TGSI_SEMANTIC_TESSOUTER:
+ return 0;
+ case TGSI_SEMANTIC_TESSINNER:
+ return 1;
+ case TGSI_SEMANTIC_PATCH:
+ assert(index < 30);
+ return 2 + index;
+
+ default:
+ assert(!"invalid semantic name");
+ return 0;
+ }
}
/**
* less than 64, so that a 64-bit bitmask of used inputs or outputs can be
* calculated.
*/
-unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
- unsigned is_varying)
-{
- switch (semantic_name) {
- case TGSI_SEMANTIC_POSITION:
- return 0;
- case TGSI_SEMANTIC_GENERIC:
- /* Since some shader stages use the the highest used IO index
- * to determine the size to allocate for inputs/outputs
- * (in LDS, tess and GS rings). GENERIC should be placed right
- * after POSITION to make that size as small as possible.
- */
- if (index < SI_MAX_IO_GENERIC)
- return 1 + index;
-
- assert(!"invalid generic index");
- return 0;
- case TGSI_SEMANTIC_FOG:
- return SI_MAX_IO_GENERIC + 1;
- case TGSI_SEMANTIC_COLOR:
- assert(index < 2);
- return SI_MAX_IO_GENERIC + 2 + index;
- case TGSI_SEMANTIC_BCOLOR:
- assert(index < 2);
- /* If it's a varying, COLOR and BCOLOR alias. */
- if (is_varying)
- return SI_MAX_IO_GENERIC + 2 + index;
- else
- return SI_MAX_IO_GENERIC + 4 + index;
- case TGSI_SEMANTIC_TEXCOORD:
- assert(index < 8);
- return SI_MAX_IO_GENERIC + 6 + index;
-
- /* These are rarely used between LS and HS or ES and GS. */
- case TGSI_SEMANTIC_CLIPDIST:
- assert(index < 2);
- return SI_MAX_IO_GENERIC + 6 + 8 + index;
- case TGSI_SEMANTIC_CLIPVERTEX:
- return SI_MAX_IO_GENERIC + 6 + 8 + 2;
- case TGSI_SEMANTIC_PSIZE:
- return SI_MAX_IO_GENERIC + 6 + 8 + 3;
-
- /* These can't be written by LS, HS, and ES. */
- case TGSI_SEMANTIC_LAYER:
- return SI_MAX_IO_GENERIC + 6 + 8 + 4;
- case TGSI_SEMANTIC_VIEWPORT_INDEX:
- return SI_MAX_IO_GENERIC + 6 + 8 + 5;
- case TGSI_SEMANTIC_PRIMID:
- STATIC_ASSERT(SI_MAX_IO_GENERIC + 6 + 8 + 6 <= 63);
- return SI_MAX_IO_GENERIC + 6 + 8 + 6;
- default:
- fprintf(stderr, "invalid semantic name = %u\n", semantic_name);
- assert(!"invalid semantic name");
- return 0;
- }
+unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index, unsigned is_varying)
+{
+ switch (semantic_name) {
+ case TGSI_SEMANTIC_POSITION:
+ return 0;
+ case TGSI_SEMANTIC_GENERIC:
+ /* Since some shader stages use the the highest used IO index
+ * to determine the size to allocate for inputs/outputs
+ * (in LDS, tess and GS rings). GENERIC should be placed right
+ * after POSITION to make that size as small as possible.
+ */
+ if (index < SI_MAX_IO_GENERIC)
+ return 1 + index;
+
+ assert(!"invalid generic index");
+ return 0;
+ case TGSI_SEMANTIC_FOG:
+ return SI_MAX_IO_GENERIC + 1;
+ case TGSI_SEMANTIC_COLOR:
+ assert(index < 2);
+ return SI_MAX_IO_GENERIC + 2 + index;
+ case TGSI_SEMANTIC_BCOLOR:
+ assert(index < 2);
+ /* If it's a varying, COLOR and BCOLOR alias. */
+ if (is_varying)
+ return SI_MAX_IO_GENERIC + 2 + index;
+ else
+ return SI_MAX_IO_GENERIC + 4 + index;
+ case TGSI_SEMANTIC_TEXCOORD:
+ assert(index < 8);
+ return SI_MAX_IO_GENERIC + 6 + index;
+
+ /* These are rarely used between LS and HS or ES and GS. */
+ case TGSI_SEMANTIC_CLIPDIST:
+ assert(index < 2);
+ return SI_MAX_IO_GENERIC + 6 + 8 + index;
+ case TGSI_SEMANTIC_CLIPVERTEX:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 2;
+ case TGSI_SEMANTIC_PSIZE:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 3;
+
+ /* These can't be written by LS, HS, and ES. */
+ case TGSI_SEMANTIC_LAYER:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 4;
+ case TGSI_SEMANTIC_VIEWPORT_INDEX:
+ return SI_MAX_IO_GENERIC + 6 + 8 + 5;
+ case TGSI_SEMANTIC_PRIMID:
+ STATIC_ASSERT(SI_MAX_IO_GENERIC + 6 + 8 + 6 <= 63);
+ return SI_MAX_IO_GENERIC + 6 + 8 + 6;
+ default:
+ fprintf(stderr, "invalid semantic name = %u\n", semantic_name);
+ assert(!"invalid semantic name");
+ return 0;
+ }
}
-/**
- * Get the value of a shader input parameter and extract a bitfield.
- */
-static LLVMValueRef unpack_llvm_param(struct si_shader_context *ctx,
- LLVMValueRef value, unsigned rshift,
- unsigned bitwidth)
+static void si_dump_streamout(struct pipe_stream_output_info *so)
{
- if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
- value = ac_to_integer(&ctx->ac, value);
-
- if (rshift)
- value = LLVMBuildLShr(ctx->ac.builder, value,
- LLVMConstInt(ctx->i32, rshift, 0), "");
-
- if (rshift + bitwidth < 32) {
- unsigned mask = (1 << bitwidth) - 1;
- value = LLVMBuildAnd(ctx->ac.builder, value,
- LLVMConstInt(ctx->i32, mask, 0), "");
- }
-
- return value;
-}
+ unsigned i;
-LLVMValueRef si_unpack_param(struct si_shader_context *ctx,
- struct ac_arg param, unsigned rshift,
- unsigned bitwidth)
-{
- LLVMValueRef value = ac_get_arg(&ctx->ac, param);
+ if (so->num_outputs)
+ fprintf(stderr, "STREAMOUT\n");
- return unpack_llvm_param(ctx, value, rshift, bitwidth);
+ for (i = 0; i < so->num_outputs; i++) {
+ unsigned mask = ((1 << so->output[i].num_components) - 1) << so->output[i].start_component;
+ fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n", i, so->output[i].output_buffer,
+ so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
+ so->output[i].register_index, mask & 1 ? "x" : "", mask & 2 ? "y" : "",
+ mask & 4 ? "z" : "", mask & 8 ? "w" : "");
+ }
}
-static LLVMValueRef unpack_sint16(struct si_shader_context *ctx,
- LLVMValueRef i32, unsigned index)
+static void declare_streamout_params(struct si_shader_context *ctx,
+ struct pipe_stream_output_info *so)
{
- assert(index <= 1);
+ if (ctx->screen->use_ngg_streamout) {
+ if (ctx->stage == MESA_SHADER_TESS_EVAL)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ return;
+ }
- if (index == 1)
- return LLVMBuildAShr(ctx->ac.builder, i32,
- LLVMConstInt(ctx->i32, 16, 0), "");
+ /* Streamout SGPRs. */
+ if (so->num_outputs) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_config);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_write_index);
+ } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ }
- return LLVMBuildSExt(ctx->ac.builder,
- LLVMBuildTrunc(ctx->ac.builder, i32,
- ctx->ac.i16, ""),
- ctx->i32, "");
-}
-
-void si_llvm_load_input_vs(
- struct si_shader_context *ctx,
- unsigned input_index,
- LLVMValueRef out[4])
-{
- const struct si_shader_info *info = &ctx->shader->selector->info;
- unsigned vs_blit_property = info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
-
- if (vs_blit_property) {
- LLVMValueRef vertex_id = ctx->abi.vertex_id;
- LLVMValueRef sel_x1 = LLVMBuildICmp(ctx->ac.builder,
- LLVMIntULE, vertex_id,
- ctx->i32_1, "");
- /* Use LLVMIntNE, because we have 3 vertices and only
- * the middle one should use y2.
- */
- LLVMValueRef sel_y1 = LLVMBuildICmp(ctx->ac.builder,
- LLVMIntNE, vertex_id,
- ctx->i32_1, "");
-
- unsigned param_vs_blit_inputs = ctx->vs_blit_inputs.arg_index;
- if (input_index == 0) {
- /* Position: */
- LLVMValueRef x1y1 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs);
- LLVMValueRef x2y2 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 1);
-
- LLVMValueRef x1 = unpack_sint16(ctx, x1y1, 0);
- LLVMValueRef y1 = unpack_sint16(ctx, x1y1, 1);
- LLVMValueRef x2 = unpack_sint16(ctx, x2y2, 0);
- LLVMValueRef y2 = unpack_sint16(ctx, x2y2, 1);
-
- LLVMValueRef x = LLVMBuildSelect(ctx->ac.builder, sel_x1,
- x1, x2, "");
- LLVMValueRef y = LLVMBuildSelect(ctx->ac.builder, sel_y1,
- y1, y2, "");
-
- out[0] = LLVMBuildSIToFP(ctx->ac.builder, x, ctx->f32, "");
- out[1] = LLVMBuildSIToFP(ctx->ac.builder, y, ctx->f32, "");
- out[2] = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 2);
- out[3] = ctx->ac.f32_1;
- return;
- }
-
- /* Color or texture coordinates: */
- assert(input_index == 1);
-
- if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
- for (int i = 0; i < 4; i++) {
- out[i] = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 3 + i);
- }
- } else {
- assert(vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD);
- LLVMValueRef x1 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 3);
- LLVMValueRef y1 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 4);
- LLVMValueRef x2 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 5);
- LLVMValueRef y2 = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 6);
-
- out[0] = LLVMBuildSelect(ctx->ac.builder, sel_x1,
- x1, x2, "");
- out[1] = LLVMBuildSelect(ctx->ac.builder, sel_y1,
- y1, y2, "");
- out[2] = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 7);
- out[3] = LLVMGetParam(ctx->main_fn,
- param_vs_blit_inputs + 8);
- }
- return;
- }
-
- unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs;
- union si_vs_fix_fetch fix_fetch;
- LLVMValueRef vb_desc;
- LLVMValueRef vertex_index;
- LLVMValueRef tmp;
-
- if (input_index < num_vbos_in_user_sgprs) {
- vb_desc = ac_get_arg(&ctx->ac, ctx->vb_descriptors[input_index]);
- } else {
- unsigned index= input_index - num_vbos_in_user_sgprs;
- vb_desc = ac_build_load_to_sgpr(&ctx->ac,
- ac_get_arg(&ctx->ac, ctx->vertex_buffers),
- LLVMConstInt(ctx->i32, index, 0));
- }
-
- vertex_index = LLVMGetParam(ctx->main_fn,
- ctx->vertex_index0.arg_index +
- input_index);
-
- /* Use the open-coded implementation for all loads of doubles and
- * of dword-sized data that needs fixups. We need to insert conversion
- * code anyway, and the amd/common code does it for us.
- *
- * Note: On LLVM <= 8, we can only open-code formats with
- * channel size >= 4 bytes.
- */
- bool opencode = ctx->shader->key.mono.vs_fetch_opencode & (1 << input_index);
- fix_fetch.bits = ctx->shader->key.mono.vs_fix_fetch[input_index].bits;
- if (opencode ||
- (fix_fetch.u.log_size == 3 && fix_fetch.u.format == AC_FETCH_FORMAT_FLOAT) ||
- (fix_fetch.u.log_size == 2)) {
- tmp = ac_build_opencoded_load_format(
- &ctx->ac, fix_fetch.u.log_size, fix_fetch.u.num_channels_m1 + 1,
- fix_fetch.u.format, fix_fetch.u.reverse, !opencode,
- vb_desc, vertex_index, ctx->ac.i32_0, ctx->ac.i32_0, 0, true);
- for (unsigned i = 0; i < 4; ++i)
- out[i] = LLVMBuildExtractElement(ctx->ac.builder, tmp, LLVMConstInt(ctx->i32, i, false), "");
- return;
- }
-
- /* Do multiple loads for special formats. */
- unsigned required_channels = util_last_bit(info->input_usage_mask[input_index]);
- LLVMValueRef fetches[4];
- unsigned num_fetches;
- unsigned fetch_stride;
- unsigned channels_per_fetch;
-
- if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2) {
- num_fetches = MIN2(required_channels, 3);
- fetch_stride = 1 << fix_fetch.u.log_size;
- channels_per_fetch = 1;
- } else {
- num_fetches = 1;
- fetch_stride = 0;
- channels_per_fetch = required_channels;
- }
-
- for (unsigned i = 0; i < num_fetches; ++i) {
- LLVMValueRef voffset = LLVMConstInt(ctx->i32, fetch_stride * i, 0);
- fetches[i] = ac_build_buffer_load_format(&ctx->ac, vb_desc, vertex_index, voffset,
- channels_per_fetch, 0, true);
- }
-
- if (num_fetches == 1 && channels_per_fetch > 1) {
- LLVMValueRef fetch = fetches[0];
- for (unsigned i = 0; i < channels_per_fetch; ++i) {
- tmp = LLVMConstInt(ctx->i32, i, false);
- fetches[i] = LLVMBuildExtractElement(
- ctx->ac.builder, fetch, tmp, "");
- }
- num_fetches = channels_per_fetch;
- channels_per_fetch = 1;
- }
-
- for (unsigned i = num_fetches; i < 4; ++i)
- fetches[i] = LLVMGetUndef(ctx->f32);
-
- if (fix_fetch.u.log_size <= 1 && fix_fetch.u.num_channels_m1 == 2 &&
- required_channels == 4) {
- if (fix_fetch.u.format == AC_FETCH_FORMAT_UINT || fix_fetch.u.format == AC_FETCH_FORMAT_SINT)
- fetches[3] = ctx->ac.i32_1;
- else
- fetches[3] = ctx->ac.f32_1;
- } else if (fix_fetch.u.log_size == 3 &&
- (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ||
- fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED ||
- fix_fetch.u.format == AC_FETCH_FORMAT_SINT) &&
- required_channels == 4) {
- /* For 2_10_10_10, the hardware returns an unsigned value;
- * convert it to a signed one.
- */
- LLVMValueRef tmp = fetches[3];
- LLVMValueRef c30 = LLVMConstInt(ctx->i32, 30, 0);
-
- /* First, recover the sign-extended signed integer value. */
- if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED)
- tmp = LLVMBuildFPToUI(ctx->ac.builder, tmp, ctx->i32, "");
- else
- tmp = ac_to_integer(&ctx->ac, tmp);
-
- /* For the integer-like cases, do a natural sign extension.
- *
- * For the SNORM case, the values are 0.0, 0.333, 0.666, 1.0
- * and happen to contain 0, 1, 2, 3 as the two LSBs of the
- * exponent.
- */
- tmp = LLVMBuildShl(ctx->ac.builder, tmp,
- fix_fetch.u.format == AC_FETCH_FORMAT_SNORM ?
- LLVMConstInt(ctx->i32, 7, 0) : c30, "");
- tmp = LLVMBuildAShr(ctx->ac.builder, tmp, c30, "");
-
- /* Convert back to the right type. */
- if (fix_fetch.u.format == AC_FETCH_FORMAT_SNORM) {
- LLVMValueRef clamp;
- LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
- tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
- clamp = LLVMBuildFCmp(ctx->ac.builder, LLVMRealULT, tmp, neg_one, "");
- tmp = LLVMBuildSelect(ctx->ac.builder, clamp, neg_one, tmp, "");
- } else if (fix_fetch.u.format == AC_FETCH_FORMAT_SSCALED) {
- tmp = LLVMBuildSIToFP(ctx->ac.builder, tmp, ctx->f32, "");
- }
-
- fetches[3] = tmp;
- }
-
- for (unsigned i = 0; i < 4; ++i)
- out[i] = ac_to_float(&ctx->ac, fetches[i]);
-}
+ /* A streamout buffer offset is loaded if the stride is non-zero. */
+ for (int i = 0; i < 4; i++) {
+ if (!so->stride[i])
+ continue;
-LLVMValueRef si_get_primitive_id(struct si_shader_context *ctx,
- unsigned swizzle)
-{
- if (swizzle > 0)
- return ctx->i32_0;
-
- switch (ctx->type) {
- case PIPE_SHADER_VERTEX:
- return ac_get_arg(&ctx->ac, ctx->vs_prim_id);
- case PIPE_SHADER_TESS_CTRL:
- return ac_get_arg(&ctx->ac, ctx->args.tcs_patch_id);
- case PIPE_SHADER_TESS_EVAL:
- return ac_get_arg(&ctx->ac, ctx->args.tes_patch_id);
- case PIPE_SHADER_GEOMETRY:
- return ac_get_arg(&ctx->ac, ctx->args.gs_prim_id);
- default:
- assert(0);
- return ctx->i32_0;
- }
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_offset[i]);
+ }
}
-static LLVMValueRef get_base_vertex(struct ac_shader_abi *abi)
+unsigned si_get_max_workgroup_size(const struct si_shader *shader)
{
- struct si_shader_context *ctx = si_shader_context_from_abi(abi);
-
- /* For non-indexed draws, the base vertex set by the driver
- * (for direct draws) or the CP (for indirect draws) is the
- * first vertex ID, but GLSL expects 0 to be returned.
- */
- LLVMValueRef vs_state = ac_get_arg(&ctx->ac,
- ctx->vs_state_bits);
- LLVMValueRef indexed;
-
- indexed = LLVMBuildLShr(ctx->ac.builder, vs_state, ctx->i32_1, "");
- indexed = LLVMBuildTrunc(ctx->ac.builder, indexed, ctx->i1, "");
-
- return LLVMBuildSelect(ctx->ac.builder, indexed,
- ac_get_arg(&ctx->ac, ctx->args.base_vertex),
- ctx->i32_0, "");
-}
+ switch (shader->selector->info.stage) {
+ case MESA_SHADER_VERTEX:
+ case MESA_SHADER_TESS_EVAL:
+ return shader->key.as_ngg ? 128 : 0;
-static LLVMValueRef get_block_size(struct ac_shader_abi *abi)
-{
- struct si_shader_context *ctx = si_shader_context_from_abi(abi);
+ case MESA_SHADER_TESS_CTRL:
+ /* Return this so that LLVM doesn't remove s_barrier
+ * instructions on chips where we use s_barrier. */
+ return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 0;
- LLVMValueRef values[3];
- LLVMValueRef result;
- unsigned i;
- unsigned *properties = ctx->shader->selector->info.properties;
+ case MESA_SHADER_GEOMETRY:
+ return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 0;
- if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
- unsigned sizes[3] = {
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
- };
+ case MESA_SHADER_COMPUTE:
+ break; /* see below */
- for (i = 0; i < 3; ++i)
- values[i] = LLVMConstInt(ctx->i32, sizes[i], 0);
+ default:
+ return 0;
+ }
- result = ac_build_gather_values(&ctx->ac, values, 3);
- } else {
- result = ac_get_arg(&ctx->ac, ctx->block_size);
- }
+ const unsigned *properties = shader->selector->info.properties;
+ unsigned max_work_group_size = properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
+ properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
+ properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
- return result;
+ if (!max_work_group_size) {
+ /* This is a variable group size compute shader,
+ * compile it for the maximum possible group size.
+ */
+ max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+ }
+ return max_work_group_size;
}
-void si_declare_compute_memory(struct si_shader_context *ctx)
+static void declare_const_and_shader_buffers(struct si_shader_context *ctx, bool assign_params)
{
- struct si_shader_selector *sel = ctx->shader->selector;
- unsigned lds_size = sel->info.properties[TGSI_PROPERTY_CS_LOCAL_SIZE];
-
- LLVMTypeRef i8p = LLVMPointerType(ctx->i8, AC_ADDR_SPACE_LDS);
- LLVMValueRef var;
-
- assert(!ctx->ac.lds);
+ enum ac_arg_type const_shader_buf_type;
- var = LLVMAddGlobalInAddressSpace(ctx->ac.module,
- LLVMArrayType(ctx->i8, lds_size),
- "compute_lds",
- AC_ADDR_SPACE_LDS);
- LLVMSetAlignment(var, 64 * 1024);
+ if (ctx->shader->selector->info.const_buffers_declared == 1 &&
+ ctx->shader->selector->info.shader_buffers_declared == 0)
+ const_shader_buf_type = AC_ARG_CONST_FLOAT_PTR;
+ else
+ const_shader_buf_type = AC_ARG_CONST_DESC_PTR;
- ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, var, i8p, "");
+ ac_add_arg(
+ &ctx->args, AC_ARG_SGPR, 1, const_shader_buf_type,
+ assign_params ? &ctx->const_and_shader_buffers : &ctx->other_const_and_shader_buffers);
}
-/* Initialize arguments for the shader export intrinsic */
-static void si_llvm_init_vs_export_args(struct si_shader_context *ctx,
- LLVMValueRef *values,
- unsigned target,
- struct ac_export_args *args)
+static void declare_samplers_and_images(struct si_shader_context *ctx, bool assign_params)
{
- args->enabled_channels = 0xf; /* writemask - default is 0xf */
- args->valid_mask = 0; /* Specify whether the EXEC mask represents the valid mask */
- args->done = 0; /* Specify whether this is the last export */
- args->target = target; /* Specify the target we are exporting */
- args->compr = false;
-
- memcpy(&args->out[0], values, sizeof(values[0]) * 4);
-}
-
-static void si_llvm_emit_clipvertex(struct si_shader_context *ctx,
- struct ac_export_args *pos, LLVMValueRef *out_elts)
-{
- unsigned reg_index;
- unsigned chan;
- unsigned const_chan;
- LLVMValueRef base_elt;
- LLVMValueRef ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
- LLVMValueRef constbuf_index = LLVMConstInt(ctx->i32,
- SI_VS_CONST_CLIP_PLANES, 0);
- LLVMValueRef const_resource = ac_build_load_to_sgpr(&ctx->ac, ptr, constbuf_index);
-
- for (reg_index = 0; reg_index < 2; reg_index ++) {
- struct ac_export_args *args = &pos[2 + reg_index];
-
- args->out[0] =
- args->out[1] =
- args->out[2] =
- args->out[3] = LLVMConstReal(ctx->f32, 0.0f);
-
- /* Compute dot products of position and user clip plane vectors */
- for (chan = 0; chan < 4; chan++) {
- for (const_chan = 0; const_chan < 4; const_chan++) {
- LLVMValueRef addr =
- LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 +
- const_chan) * 4, 0);
- base_elt = si_buffer_load_const(ctx, const_resource,
- addr);
- args->out[chan] = ac_build_fmad(&ctx->ac, base_elt,
- out_elts[const_chan], args->out[chan]);
- }
- }
-
- args->enabled_channels = 0xf;
- args->valid_mask = 0;
- args->done = 0;
- args->target = V_008DFC_SQ_EXP_POS + 2 + reg_index;
- args->compr = 0;
- }
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
+ assign_params ? &ctx->samplers_and_images : &ctx->other_samplers_and_images);
}
-static void si_dump_streamout(struct pipe_stream_output_info *so)
+static void declare_per_stage_desc_pointers(struct si_shader_context *ctx, bool assign_params)
{
- unsigned i;
-
- if (so->num_outputs)
- fprintf(stderr, "STREAMOUT\n");
-
- for (i = 0; i < so->num_outputs; i++) {
- unsigned mask = ((1 << so->output[i].num_components) - 1) <<
- so->output[i].start_component;
- fprintf(stderr, " %i: BUF%i[%i..%i] <- OUT[%i].%s%s%s%s\n",
- i, so->output[i].output_buffer,
- so->output[i].dst_offset, so->output[i].dst_offset + so->output[i].num_components - 1,
- so->output[i].register_index,
- mask & 1 ? "x" : "",
- mask & 2 ? "y" : "",
- mask & 4 ? "z" : "",
- mask & 8 ? "w" : "");
- }
+ declare_const_and_shader_buffers(ctx, assign_params);
+ declare_samplers_and_images(ctx, assign_params);
}
-void si_emit_streamout_output(struct si_shader_context *ctx,
- LLVMValueRef const *so_buffers,
- LLVMValueRef const *so_write_offsets,
- struct pipe_stream_output *stream_out,
- struct si_shader_output_values *shader_out)
-{
- unsigned buf_idx = stream_out->output_buffer;
- unsigned start = stream_out->start_component;
- unsigned num_comps = stream_out->num_components;
- LLVMValueRef out[4];
-
- assert(num_comps && num_comps <= 4);
- if (!num_comps || num_comps > 4)
- return;
-
- /* Load the output as int. */
- for (int j = 0; j < num_comps; j++) {
- assert(stream_out->stream == shader_out->vertex_stream[start + j]);
-
- out[j] = ac_to_integer(&ctx->ac, shader_out->values[start + j]);
- }
-
- /* Pack the output. */
- LLVMValueRef vdata = NULL;
-
- switch (num_comps) {
- case 1: /* as i32 */
- vdata = out[0];
- break;
- case 2: /* as v2i32 */
- case 3: /* as v3i32 */
- if (ac_has_vec3_support(ctx->screen->info.chip_class, false)) {
- vdata = ac_build_gather_values(&ctx->ac, out, num_comps);
- break;
- }
- /* as v4i32 (aligned to 4) */
- out[3] = LLVMGetUndef(ctx->i32);
- /* fall through */
- case 4: /* as v4i32 */
- vdata = ac_build_gather_values(&ctx->ac, out, util_next_power_of_two(num_comps));
- break;
- }
-
- ac_build_buffer_store_dword(&ctx->ac, so_buffers[buf_idx],
- vdata, num_comps,
- so_write_offsets[buf_idx],
- ctx->i32_0,
- stream_out->dst_offset * 4, ac_glc | ac_slc);
-}
-
-/**
- * Write streamout data to buffers for vertex stream @p stream (different
- * vertex streams can occur for GS copy shaders).
- */
-void si_llvm_emit_streamout(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput, unsigned stream)
-{
- struct si_shader_selector *sel = ctx->shader->selector;
- struct pipe_stream_output_info *so = &sel->so;
- LLVMBuilderRef builder = ctx->ac.builder;
- int i;
-
- /* Get bits [22:16], i.e. (so_param >> 16) & 127; */
- LLVMValueRef so_vtx_count =
- si_unpack_param(ctx, ctx->streamout_config, 16, 7);
-
- LLVMValueRef tid = ac_get_thread_id(&ctx->ac);
-
- /* can_emit = tid < so_vtx_count; */
- LLVMValueRef can_emit =
- LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
-
- /* Emit the streamout code conditionally. This actually avoids
- * out-of-bounds buffer access. The hw tells us via the SGPR
- * (so_vtx_count) which threads are allowed to emit streamout data. */
- ac_build_ifcc(&ctx->ac, can_emit, 6501);
- {
- /* The buffer offset is computed as follows:
- * ByteOffset = streamout_offset[buffer_id]*4 +
- * (streamout_write_index + thread_id)*stride[buffer_id] +
- * attrib_offset
- */
-
- LLVMValueRef so_write_index =
- ac_get_arg(&ctx->ac,
- ctx->streamout_write_index);
-
- /* Compute (streamout_write_index + thread_id). */
- so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
-
- /* Load the descriptor and compute the write offset for each
- * enabled buffer. */
- LLVMValueRef so_write_offset[4] = {};
- LLVMValueRef so_buffers[4];
- LLVMValueRef buf_ptr = ac_get_arg(&ctx->ac,
- ctx->rw_buffers);
-
- for (i = 0; i < 4; i++) {
- if (!so->stride[i])
- continue;
-
- LLVMValueRef offset = LLVMConstInt(ctx->i32,
- SI_VS_STREAMOUT_BUF0 + i, 0);
-
- so_buffers[i] = ac_build_load_to_sgpr(&ctx->ac, buf_ptr, offset);
-
- LLVMValueRef so_offset = ac_get_arg(&ctx->ac,
- ctx->streamout_offset[i]);
- so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
-
- so_write_offset[i] = ac_build_imad(&ctx->ac, so_write_index,
- LLVMConstInt(ctx->i32, so->stride[i]*4, 0),
- so_offset);
- }
-
- /* Write streamout data. */
- for (i = 0; i < so->num_outputs; i++) {
- unsigned reg = so->output[i].register_index;
-
- if (reg >= noutput)
- continue;
-
- if (stream != so->output[i].stream)
- continue;
-
- si_emit_streamout_output(ctx, so_buffers, so_write_offset,
- &so->output[i], &outputs[reg]);
- }
- }
- ac_build_endif(&ctx->ac, 6501);
-}
-
-static void si_export_param(struct si_shader_context *ctx, unsigned index,
- LLVMValueRef *values)
-{
- struct ac_export_args args;
-
- si_llvm_init_vs_export_args(ctx, values,
- V_008DFC_SQ_EXP_PARAM + index, &args);
- ac_build_export(&ctx->ac, &args);
-}
-
-static void si_build_param_exports(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
-{
- struct si_shader *shader = ctx->shader;
- unsigned param_count = 0;
-
- for (unsigned i = 0; i < noutput; i++) {
- unsigned semantic_name = outputs[i].semantic_name;
- unsigned semantic_index = outputs[i].semantic_index;
-
- if (outputs[i].vertex_stream[0] != 0 &&
- outputs[i].vertex_stream[1] != 0 &&
- outputs[i].vertex_stream[2] != 0 &&
- outputs[i].vertex_stream[3] != 0)
- continue;
-
- switch (semantic_name) {
- case TGSI_SEMANTIC_LAYER:
- case TGSI_SEMANTIC_VIEWPORT_INDEX:
- case TGSI_SEMANTIC_CLIPDIST:
- case TGSI_SEMANTIC_COLOR:
- case TGSI_SEMANTIC_BCOLOR:
- case TGSI_SEMANTIC_PRIMID:
- case TGSI_SEMANTIC_FOG:
- case TGSI_SEMANTIC_TEXCOORD:
- case TGSI_SEMANTIC_GENERIC:
- break;
- default:
- continue;
- }
-
- if ((semantic_name != TGSI_SEMANTIC_GENERIC ||
- semantic_index < SI_MAX_IO_GENERIC) &&
- shader->key.opt.kill_outputs &
- (1ull << si_shader_io_get_unique_index(semantic_name,
- semantic_index, true)))
- continue;
-
- si_export_param(ctx, param_count, outputs[i].values);
-
- assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
- shader->info.vs_output_param_offset[i] = param_count++;
- }
-
- shader->info.nr_param_exports = param_count;
-}
-
-/**
- * Vertex color clamping.
- *
- * This uses a state constant loaded in a user data SGPR and
- * an IF statement is added that clamps all colors if the constant
- * is true.
- */
-static void si_vertex_color_clamping(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
-{
- LLVMValueRef addr[SI_MAX_VS_OUTPUTS][4];
- bool has_colors = false;
-
- /* Store original colors to alloca variables. */
- for (unsigned i = 0; i < noutput; i++) {
- if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
- outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
- continue;
-
- for (unsigned j = 0; j < 4; j++) {
- addr[i][j] = ac_build_alloca_undef(&ctx->ac, ctx->f32, "");
- LLVMBuildStore(ctx->ac.builder, outputs[i].values[j], addr[i][j]);
- }
- has_colors = true;
- }
-
- if (!has_colors)
- return;
-
- /* The state is in the first bit of the user SGPR. */
- LLVMValueRef cond = ac_get_arg(&ctx->ac, ctx->vs_state_bits);
- cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->i1, "");
-
- ac_build_ifcc(&ctx->ac, cond, 6502);
-
- /* Store clamped colors to alloca variables within the conditional block. */
- for (unsigned i = 0; i < noutput; i++) {
- if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
- outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
- continue;
-
- for (unsigned j = 0; j < 4; j++) {
- LLVMBuildStore(ctx->ac.builder,
- ac_build_clamp(&ctx->ac, outputs[i].values[j]),
- addr[i][j]);
- }
- }
- ac_build_endif(&ctx->ac, 6502);
-
- /* Load clamped colors */
- for (unsigned i = 0; i < noutput; i++) {
- if (outputs[i].semantic_name != TGSI_SEMANTIC_COLOR &&
- outputs[i].semantic_name != TGSI_SEMANTIC_BCOLOR)
- continue;
-
- for (unsigned j = 0; j < 4; j++) {
- outputs[i].values[j] =
- LLVMBuildLoad(ctx->ac.builder, addr[i][j], "");
- }
- }
-}
-
-/* Generate export instructions for hardware VS shader stage or NGG GS stage
- * (position and parameter data only).
- */
-void si_llvm_export_vs(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
-{
- struct si_shader *shader = ctx->shader;
- struct ac_export_args pos_args[4] = {};
- LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
- unsigned pos_idx;
- int i;
-
- si_vertex_color_clamping(ctx, outputs, noutput);
-
- /* Build position exports. */
- for (i = 0; i < noutput; i++) {
- switch (outputs[i].semantic_name) {
- case TGSI_SEMANTIC_POSITION:
- si_llvm_init_vs_export_args(ctx, outputs[i].values,
- V_008DFC_SQ_EXP_POS, &pos_args[0]);
- break;
- case TGSI_SEMANTIC_PSIZE:
- psize_value = outputs[i].values[0];
- break;
- case TGSI_SEMANTIC_LAYER:
- layer_value = outputs[i].values[0];
- break;
- case TGSI_SEMANTIC_VIEWPORT_INDEX:
- viewport_index_value = outputs[i].values[0];
- break;
- case TGSI_SEMANTIC_EDGEFLAG:
- edgeflag_value = outputs[i].values[0];
- break;
- case TGSI_SEMANTIC_CLIPDIST:
- if (!shader->key.opt.clip_disable) {
- unsigned index = 2 + outputs[i].semantic_index;
- si_llvm_init_vs_export_args(ctx, outputs[i].values,
- V_008DFC_SQ_EXP_POS + index,
- &pos_args[index]);
- }
- break;
- case TGSI_SEMANTIC_CLIPVERTEX:
- if (!shader->key.opt.clip_disable) {
- si_llvm_emit_clipvertex(ctx, pos_args,
- outputs[i].values);
- }
- break;
- }
- }
-
- /* We need to add the position output manually if it's missing. */
- if (!pos_args[0].out[0]) {
- pos_args[0].enabled_channels = 0xf; /* writemask */
- pos_args[0].valid_mask = 0; /* EXEC mask */
- pos_args[0].done = 0; /* last export? */
- pos_args[0].target = V_008DFC_SQ_EXP_POS;
- pos_args[0].compr = 0; /* COMPR flag */
- pos_args[0].out[0] = ctx->ac.f32_0; /* X */
- pos_args[0].out[1] = ctx->ac.f32_0; /* Y */
- pos_args[0].out[2] = ctx->ac.f32_0; /* Z */
- pos_args[0].out[3] = ctx->ac.f32_1; /* W */
- }
-
- bool pos_writes_edgeflag = shader->selector->info.writes_edgeflag &&
- !shader->key.as_ngg;
-
- /* Write the misc vector (point size, edgeflag, layer, viewport). */
- if (shader->selector->info.writes_psize ||
- pos_writes_edgeflag ||
- shader->selector->info.writes_viewport_index ||
- shader->selector->info.writes_layer) {
- pos_args[1].enabled_channels = shader->selector->info.writes_psize |
- (pos_writes_edgeflag << 1) |
- (shader->selector->info.writes_layer << 2);
-
- pos_args[1].valid_mask = 0; /* EXEC mask */
- pos_args[1].done = 0; /* last export? */
- pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
- pos_args[1].compr = 0; /* COMPR flag */
- pos_args[1].out[0] = ctx->ac.f32_0; /* X */
- pos_args[1].out[1] = ctx->ac.f32_0; /* Y */
- pos_args[1].out[2] = ctx->ac.f32_0; /* Z */
- pos_args[1].out[3] = ctx->ac.f32_0; /* W */
-
- if (shader->selector->info.writes_psize)
- pos_args[1].out[0] = psize_value;
-
- if (pos_writes_edgeflag) {
- /* The output is a float, but the hw expects an integer
- * with the first bit containing the edge flag. */
- edgeflag_value = LLVMBuildFPToUI(ctx->ac.builder,
- edgeflag_value,
- ctx->i32, "");
- edgeflag_value = ac_build_umin(&ctx->ac,
- edgeflag_value,
- ctx->i32_1);
-
- /* The LLVM intrinsic expects a float. */
- pos_args[1].out[1] = ac_to_float(&ctx->ac, edgeflag_value);
- }
-
- if (ctx->screen->info.chip_class >= GFX9) {
- /* GFX9 has the layer in out.z[10:0] and the viewport
- * index in out.z[19:16].
- */
- if (shader->selector->info.writes_layer)
- pos_args[1].out[2] = layer_value;
-
- if (shader->selector->info.writes_viewport_index) {
- LLVMValueRef v = viewport_index_value;
-
- v = ac_to_integer(&ctx->ac, v);
- v = LLVMBuildShl(ctx->ac.builder, v,
- LLVMConstInt(ctx->i32, 16, 0), "");
- v = LLVMBuildOr(ctx->ac.builder, v,
- ac_to_integer(&ctx->ac, pos_args[1].out[2]), "");
- pos_args[1].out[2] = ac_to_float(&ctx->ac, v);
- pos_args[1].enabled_channels |= 1 << 2;
- }
- } else {
- if (shader->selector->info.writes_layer)
- pos_args[1].out[2] = layer_value;
-
- if (shader->selector->info.writes_viewport_index) {
- pos_args[1].out[3] = viewport_index_value;
- pos_args[1].enabled_channels |= 1 << 3;
- }
- }
- }
-
- for (i = 0; i < 4; i++)
- if (pos_args[i].out[0])
- shader->info.nr_pos_exports++;
-
- /* Navi10-14 skip POS0 exports if EXEC=0 and DONE=0, causing a hang.
- * Setting valid_mask=1 prevents it and has no other effect.
- */
- if (ctx->screen->info.family == CHIP_NAVI10 ||
- ctx->screen->info.family == CHIP_NAVI12 ||
- ctx->screen->info.family == CHIP_NAVI14)
- pos_args[0].valid_mask = 1;
-
- pos_idx = 0;
- for (i = 0; i < 4; i++) {
- if (!pos_args[i].out[0])
- continue;
-
- /* Specify the target we are exporting */
- pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;
-
- if (pos_idx == shader->info.nr_pos_exports)
- /* Specify that this is the last export */
- pos_args[i].done = 1;
-
- ac_build_export(&ctx->ac, &pos_args[i]);
- }
-
- /* Build parameter exports. */
- si_build_param_exports(ctx, outputs, noutput);
-}
-
-static void si_llvm_emit_vs_epilogue(struct ac_shader_abi *abi,
- unsigned max_outputs,
- LLVMValueRef *addrs)
+static void declare_global_desc_pointers(struct si_shader_context *ctx)
{
- struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- struct si_shader_info *info = &ctx->shader->selector->info;
- struct si_shader_output_values *outputs = NULL;
- int i,j;
-
- assert(!ctx->shader->is_gs_copy_shader);
- assert(info->num_outputs <= max_outputs);
-
- outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
-
- for (i = 0; i < info->num_outputs; i++) {
- outputs[i].semantic_name = info->output_semantic_name[i];
- outputs[i].semantic_index = info->output_semantic_index[i];
-
- for (j = 0; j < 4; j++) {
- outputs[i].values[j] =
- LLVMBuildLoad(ctx->ac.builder,
- addrs[4 * i + j],
- "");
- outputs[i].vertex_stream[j] =
- (info->output_streams[i] >> (2 * j)) & 3;
- }
- }
-
- if (!ctx->screen->use_ngg_streamout &&
- ctx->shader->selector->so.num_outputs)
- si_llvm_emit_streamout(ctx, outputs, i, 0);
-
- /* Export PrimitiveID. */
- if (ctx->shader->key.mono.u.vs_export_prim_id) {
- outputs[i].semantic_name = TGSI_SEMANTIC_PRIMID;
- outputs[i].semantic_index = 0;
- outputs[i].values[0] = ac_to_float(&ctx->ac, si_get_primitive_id(ctx, 0));
- for (j = 1; j < 4; j++)
- outputs[i].values[j] = LLVMConstReal(ctx->f32, 0);
-
- memset(outputs[i].vertex_stream, 0,
- sizeof(outputs[i].vertex_stream));
- i++;
- }
-
- si_llvm_export_vs(ctx, outputs, i);
- FREE(outputs);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->rw_buffers);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
+ &ctx->bindless_samplers_and_images);
}
-static void si_llvm_emit_prim_discard_cs_epilogue(struct ac_shader_abi *abi,
- unsigned max_outputs,
- LLVMValueRef *addrs)
+static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx)
{
- struct si_shader_context *ctx = si_shader_context_from_abi(abi);
- struct si_shader_info *info = &ctx->shader->selector->info;
- LLVMValueRef pos[4] = {};
-
- assert(info->num_outputs <= max_outputs);
-
- for (unsigned i = 0; i < info->num_outputs; i++) {
- if (info->output_semantic_name[i] != TGSI_SEMANTIC_POSITION)
- continue;
-
- for (unsigned chan = 0; chan < 4; chan++)
- pos[chan] = LLVMBuildLoad(ctx->ac.builder, addrs[4 * i + chan], "");
- break;
- }
- assert(pos[0] != NULL);
-
- /* Return the position output. */
- LLVMValueRef ret = ctx->return_value;
- for (unsigned chan = 0; chan < 4; chan++)
- ret = LLVMBuildInsertValue(ctx->ac.builder, ret, pos[chan], chan, "");
- ctx->return_value = ret;
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ if (!ctx->shader->is_gs_copy_shader) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.base_vertex);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.start_instance);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.draw_id);
+ }
}
-static void declare_streamout_params(struct si_shader_context *ctx,
- struct pipe_stream_output_info *so)
+static void declare_vb_descriptor_input_sgprs(struct si_shader_context *ctx)
{
- if (ctx->screen->use_ngg_streamout) {
- if (ctx->type == PIPE_SHADER_TESS_EVAL)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
- return;
- }
-
- /* Streamout SGPRs. */
- if (so->num_outputs) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_config);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_write_index);
- } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
- }
-
- /* A streamout buffer offset is loaded if the stride is non-zero. */
- for (int i = 0; i < 4; i++) {
- if (!so->stride[i])
- continue;
-
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->streamout_offset[i]);
- }
-}
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->vertex_buffers);
+
+ unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs;
+ if (num_vbos_in_user_sgprs) {
+ unsigned user_sgprs = ctx->args.num_sgprs_used;
+
+ if (si_is_merged_shader(ctx->shader))
+ user_sgprs -= 8;
+ assert(user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST);
+
+ /* Declare unused SGPRs to align VB descriptors to 4 SGPRs (hw requirement). */
+ for (unsigned i = user_sgprs; i < SI_SGPR_VS_VB_DESCRIPTOR_FIRST; i++)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+
+ assert(num_vbos_in_user_sgprs <= ARRAY_SIZE(ctx->vb_descriptors));
+ for (unsigned i = 0; i < num_vbos_in_user_sgprs; i++)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 4, AC_ARG_INT, &ctx->vb_descriptors[i]);
+ }
+}
+
+static void declare_vs_input_vgprs(struct si_shader_context *ctx, unsigned *num_prolog_vgprs,
+ bool ngg_cull_shader)
+{
+ struct si_shader *shader = ctx->shader;
+
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.vertex_id);
+ if (shader->key.as_ls) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->rel_auto_id);
+ if (ctx->screen->info.chip_class >= GFX10) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ } else {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
+ }
+ } else if (ctx->screen->info.chip_class >= GFX10) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
+ &ctx->vs_prim_id); /* user vgpr or PrimID (legacy) */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ } else {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
+ }
+
+ if (!shader->is_gs_copy_shader) {
+ if (shader->key.opt.ngg_culling && !ngg_cull_shader) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->ngg_old_thread_id);
+ }
+
+ /* Vertex load indices. */
+ if (shader->selector->info.num_inputs) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vertex_index0);
+ for (unsigned i = 1; i < shader->selector->info.num_inputs; i++)
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
+ }
+ *num_prolog_vgprs += shader->selector->info.num_inputs;
+ }
+}
+
+static void declare_vs_blit_inputs(struct si_shader_context *ctx, unsigned vs_blit_property)
+{
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_blit_inputs); /* i16 x1, y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* i16 x1, y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* depth */
+
+ if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color0 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color3 */
+ } else if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y1 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y2 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.z */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.w */
+ }
+}
+
+static void declare_tes_input_vgprs(struct si_shader_context *ctx, bool ngg_cull_shader)
+{
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_u);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_v);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->tes_rel_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tes_patch_id);
+
+ if (ctx->shader->key.opt.ngg_culling && !ngg_cull_shader) {
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->ngg_old_thread_id);
+ }
+}
+
+enum
+{
+ /* Convenient merged shader definitions. */
+ SI_SHADER_MERGED_VERTEX_TESSCTRL = MESA_ALL_SHADER_STAGES,
+ SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY,
+};
-static unsigned si_get_max_workgroup_size(const struct si_shader *shader)
-{
- switch (shader->selector->type) {
- case PIPE_SHADER_VERTEX:
- case PIPE_SHADER_TESS_EVAL:
- return shader->key.as_ngg ? 128 : 0;
-
- case PIPE_SHADER_TESS_CTRL:
- /* Return this so that LLVM doesn't remove s_barrier
- * instructions on chips where we use s_barrier. */
- return shader->selector->screen->info.chip_class >= GFX7 ? 128 : 0;
-
- case PIPE_SHADER_GEOMETRY:
- return shader->selector->screen->info.chip_class >= GFX9 ? 128 : 0;
-
- case PIPE_SHADER_COMPUTE:
- break; /* see below */
-
- default:
- return 0;
- }
-
- const unsigned *properties = shader->selector->info.properties;
- unsigned max_work_group_size =
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
- properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
-
- if (!max_work_group_size) {
- /* This is a variable group size compute shader,
- * compile it for the maximum possible group size.
- */
- max_work_group_size = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
- }
- return max_work_group_size;
+void si_add_arg_checked(struct ac_shader_args *args, enum ac_arg_regfile file, unsigned registers,
+ enum ac_arg_type type, struct ac_arg *arg, unsigned idx)
+{
+ assert(args->arg_count == idx);
+ ac_add_arg(args, file, registers, type, arg);
+}
+
+void si_create_function(struct si_shader_context *ctx, bool ngg_cull_shader)
+{
+ struct si_shader *shader = ctx->shader;
+ LLVMTypeRef returns[AC_MAX_ARGS];
+ unsigned i, num_return_sgprs;
+ unsigned num_returns = 0;
+ unsigned num_prolog_vgprs = 0;
+ unsigned stage = ctx->stage;
+ unsigned vs_blit_property = shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
+
+ memset(&ctx->args, 0, sizeof(ctx->args));
+
+ /* Set MERGED shaders. */
+ if (ctx->screen->info.chip_class >= GFX9) {
+ if (shader->key.as_ls || stage == MESA_SHADER_TESS_CTRL)
+ stage = SI_SHADER_MERGED_VERTEX_TESSCTRL; /* LS or HS */
+ else if (shader->key.as_es || shader->key.as_ngg || stage == MESA_SHADER_GEOMETRY)
+ stage = SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY;
+ }
+
+ switch (stage) {
+ case MESA_SHADER_VERTEX:
+ declare_global_desc_pointers(ctx);
+
+ if (vs_blit_property) {
+ declare_vs_blit_inputs(ctx, vs_blit_property);
+
+ /* VGPRs */
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs, ngg_cull_shader);
+ break;
+ }
+
+ declare_per_stage_desc_pointers(ctx, true);
+ declare_vs_specific_input_sgprs(ctx);
+ if (!shader->is_gs_copy_shader)
+ declare_vb_descriptor_input_sgprs(ctx);
+
+ if (shader->key.as_es) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->es2gs_offset);
+ } else if (shader->key.as_ls) {
+ /* no extra parameters */
+ } else {
+ /* The locations of the other parameters are assigned dynamically. */
+ declare_streamout_params(ctx, &shader->selector->so);
+ }
+
+ /* VGPRs */
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs, ngg_cull_shader);
+
+ /* Return values */
+ if (shader->key.opt.vs_as_prim_discard_cs) {
+ for (i = 0; i < 4; i++)
+ returns[num_returns++] = ctx->ac.f32; /* VGPRs */
+ }
+ break;
+
+ case MESA_SHADER_TESS_CTRL: /* GFX6-GFX8 */
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
+
+ /* VGPRs */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
+
+ /* param_tcs_offchip_offset and param_tcs_factor_offset are
+ * placed after the user SGPRs.
+ */
+ for (i = 0; i < GFX6_TCS_NUM_USER_SGPR + 2; i++)
+ returns[num_returns++] = ctx->ac.i32; /* SGPRs */
+ for (i = 0; i < 11; i++)
+ returns[num_returns++] = ctx->ac.f32; /* VGPRs */
+ break;
+
+ case SI_SHADER_MERGED_VERTEX_TESSCTRL:
+ /* Merged stages have 8 system SGPRs at the beginning. */
+ /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
+ declare_per_stage_desc_pointers(ctx, ctx->stage == MESA_SHADER_TESS_CTRL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, ctx->stage == MESA_SHADER_VERTEX);
+ declare_vs_specific_input_sgprs(ctx);
+
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
+ declare_vb_descriptor_input_sgprs(ctx);
+
+ /* VGPRs (first TCS, then VS) */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs, ngg_cull_shader);
+
+ /* LS return values are inputs to the TCS main shader part. */
+ for (i = 0; i < 8 + GFX9_TCS_NUM_USER_SGPR; i++)
+ returns[num_returns++] = ctx->ac.i32; /* SGPRs */
+ for (i = 0; i < 2; i++)
+ returns[num_returns++] = ctx->ac.f32; /* VGPRs */
+ } else {
+ /* TCS return values are inputs to the TCS epilog.
+ *
+ * param_tcs_offchip_offset, param_tcs_factor_offset,
+ * param_tcs_offchip_layout, and param_rw_buffers
+ * should be passed to the epilog.
+ */
+ for (i = 0; i <= 8 + GFX9_SGPR_TCS_OUT_LAYOUT; i++)
+ returns[num_returns++] = ctx->ac.i32; /* SGPRs */
+ for (i = 0; i < 11; i++)
+ returns[num_returns++] = ctx->ac.f32; /* VGPRs */
+ }
+ break;
+
+ case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
+ /* Merged stages have 8 system SGPRs at the beginning. */
+ /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
+ declare_per_stage_desc_pointers(ctx, ctx->stage == MESA_SHADER_GEOMETRY);
+
+ if (ctx->shader->key.as_ngg)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_tg_info);
+ else
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
+
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
+ &ctx->small_prim_cull_info); /* SPI_SHADER_PGM_LO_GS << 8 */
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
+ NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
+
+ declare_global_desc_pointers(ctx);
+ if (ctx->stage != MESA_SHADER_VERTEX || !vs_blit_property) {
+ declare_per_stage_desc_pointers(
+ ctx, (ctx->stage == MESA_SHADER_VERTEX || ctx->stage == MESA_SHADER_TESS_EVAL));
+ }
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ if (vs_blit_property)
+ declare_vs_blit_inputs(ctx, vs_blit_property);
+ else
+ declare_vs_specific_input_sgprs(ctx);
+ } else {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
+ /* Declare as many input SGPRs as the VS has. */
+ }
+
+ if (ctx->stage == MESA_SHADER_VERTEX)
+ declare_vb_descriptor_input_sgprs(ctx);
+
+ /* VGPRs (first GS, then VS/TES) */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx01_offset);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx23_offset);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx45_offset);
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ declare_vs_input_vgprs(ctx, &num_prolog_vgprs, ngg_cull_shader);
+ } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
+ declare_tes_input_vgprs(ctx, ngg_cull_shader);
+ }
+
+ if ((ctx->shader->key.as_es || ngg_cull_shader) &&
+ (ctx->stage == MESA_SHADER_VERTEX || ctx->stage == MESA_SHADER_TESS_EVAL)) {
+ unsigned num_user_sgprs, num_vgprs;
+
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ /* For the NGG cull shader, add 1 SGPR to hold
+ * the vertex buffer pointer.
+ */
+ num_user_sgprs = GFX9_VSGS_NUM_USER_SGPR + ngg_cull_shader;
+
+ if (ngg_cull_shader && shader->selector->num_vbos_in_user_sgprs) {
+ assert(num_user_sgprs <= 8 + SI_SGPR_VS_VB_DESCRIPTOR_FIRST);
+ num_user_sgprs =
+ SI_SGPR_VS_VB_DESCRIPTOR_FIRST + shader->selector->num_vbos_in_user_sgprs * 4;
+ }
+ } else {
+ num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
+ }
+
+ /* The NGG cull shader has to return all 9 VGPRs + the old thread ID.
+ *
+ * The normal merged ESGS shader only has to return the 5 VGPRs
+ * for the GS stage.
+ */
+ num_vgprs = ngg_cull_shader ? 10 : 5;
+
+ /* ES return values are inputs to GS. */
+ for (i = 0; i < 8 + num_user_sgprs; i++)
+ returns[num_returns++] = ctx->ac.i32; /* SGPRs */
+ for (i = 0; i < num_vgprs; i++)
+ returns[num_returns++] = ctx->ac.f32; /* VGPRs */
+ }
+ break;
+
+ case MESA_SHADER_TESS_EVAL:
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
+
+ if (shader->key.as_es) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->es2gs_offset);
+ } else {
+ declare_streamout_params(ctx, &shader->selector->so);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
+ }
+
+ /* VGPRs */
+ declare_tes_input_vgprs(ctx, ngg_cull_shader);
+ break;
+
+ case MESA_SHADER_GEOMETRY:
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_wave_id);
+
+ /* VGPRs */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[0]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[1]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[2]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[3]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[4]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[5]);
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
+ break;
+
+ case MESA_SHADER_FRAGMENT:
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL, SI_PARAM_ALPHA_REF);
+ si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.prim_mask,
+ SI_PARAM_PRIM_MASK);
+
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_sample,
+ SI_PARAM_PERSP_SAMPLE);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_center,
+ SI_PARAM_PERSP_CENTER);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_centroid,
+ SI_PARAM_PERSP_CENTROID);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT, NULL, SI_PARAM_PERSP_PULL_MODEL);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.linear_sample,
+ SI_PARAM_LINEAR_SAMPLE);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.linear_center,
+ SI_PARAM_LINEAR_CENTER);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.linear_centroid,
+ SI_PARAM_LINEAR_CENTROID);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_FLOAT, NULL, SI_PARAM_LINE_STIPPLE_TEX);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.frag_pos[0],
+ SI_PARAM_POS_X_FLOAT);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.frag_pos[1],
+ SI_PARAM_POS_Y_FLOAT);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.frag_pos[2],
+ SI_PARAM_POS_Z_FLOAT);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.frag_pos[3],
+ SI_PARAM_POS_W_FLOAT);
+ shader->info.face_vgpr_index = ctx->args.num_vgprs_used;
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.front_face,
+ SI_PARAM_FRONT_FACE);
+ shader->info.ancillary_vgpr_index = ctx->args.num_vgprs_used;
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.ancillary,
+ SI_PARAM_ANCILLARY);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->args.sample_coverage,
+ SI_PARAM_SAMPLE_COVERAGE);
+ si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->pos_fixed_pt,
+ SI_PARAM_POS_FIXED_PT);
+
+ /* Color inputs from the prolog. */
+ if (shader->selector->info.colors_read) {
+ unsigned num_color_elements = util_bitcount(shader->selector->info.colors_read);
+
+ for (i = 0; i < num_color_elements; i++)
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
+
+ num_prolog_vgprs += num_color_elements;
+ }
+
+ /* Outputs for the epilog. */
+ num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
+ num_returns = num_return_sgprs + util_bitcount(shader->selector->info.colors_written) * 4 +
+ shader->selector->info.writes_z + shader->selector->info.writes_stencil +
+ shader->selector->info.writes_samplemask + 1 /* SampleMaskIn */;
+
+ num_returns = MAX2(num_returns, num_return_sgprs + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
+
+ for (i = 0; i < num_return_sgprs; i++)
+ returns[i] = ctx->ac.i32;
+ for (; i < num_returns; i++)
+ returns[i] = ctx->ac.f32;
+ break;
+
+ case MESA_SHADER_COMPUTE:
+ declare_global_desc_pointers(ctx);
+ declare_per_stage_desc_pointers(ctx, true);
+ if (shader->selector->info.uses_grid_size)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT, &ctx->args.num_work_groups);
+ if (shader->selector->info.uses_block_size &&
+ shader->selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT, &ctx->block_size);
+
+ unsigned cs_user_data_dwords =
+ shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
+ if (cs_user_data_dwords) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, cs_user_data_dwords, AC_ARG_INT, &ctx->cs_user_data);
+ }
+
+ /* Some descriptors can be in user SGPRs. */
+ /* Shader buffers in user SGPRs. */
+ for (unsigned i = 0; i < shader->selector->cs_num_shaderbufs_in_user_sgprs; i++) {
+ while (ctx->args.num_sgprs_used % 4 != 0)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 4, AC_ARG_INT, &ctx->cs_shaderbuf[i]);
+ }
+ /* Images in user SGPRs. */
+ for (unsigned i = 0; i < shader->selector->cs_num_images_in_user_sgprs; i++) {
+ unsigned num_sgprs = shader->selector->info.image_buffers & (1 << i) ? 4 : 8;
+
+ while (ctx->args.num_sgprs_used % num_sgprs != 0)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
+
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, num_sgprs, AC_ARG_INT, &ctx->cs_image[i]);
+ }
+
+ /* Hardware SGPRs. */
+ for (i = 0; i < 3; i++) {
+ if (shader->selector->info.uses_block_id[i]) {
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.workgroup_ids[i]);
+ }
+ }
+ if (shader->selector->info.uses_subgroup_info)
+ ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.tg_size);
+
+ /* Hardware VGPRs. */
+ ac_add_arg(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT, &ctx->args.local_invocation_ids);
+ break;
+ default:
+ assert(0 && "unimplemented shader");
+ return;
+ }
+
+ si_llvm_create_func(ctx, ngg_cull_shader ? "ngg_cull_main" : "main", returns, num_returns,
+ si_get_max_workgroup_size(shader));
+
+ /* Reserve register locations for VGPR inputs the PS prolog may need. */
+ if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->shader->is_monolithic) {
+ ac_llvm_add_target_dep_function_attr(
+ ctx->main_fn, "InitialPSInputAddr",
+ S_0286D0_PERSP_SAMPLE_ENA(1) | S_0286D0_PERSP_CENTER_ENA(1) |
+ S_0286D0_PERSP_CENTROID_ENA(1) | S_0286D0_LINEAR_SAMPLE_ENA(1) |
+ S_0286D0_LINEAR_CENTER_ENA(1) | S_0286D0_LINEAR_CENTROID_ENA(1) |
+ S_0286D0_FRONT_FACE_ENA(1) | S_0286D0_ANCILLARY_ENA(1) | S_0286D0_POS_FIXED_PT_ENA(1));
+ }
+
+ shader->info.num_input_sgprs = ctx->args.num_sgprs_used;
+ shader->info.num_input_vgprs = ctx->args.num_vgprs_used;
+
+ assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
+ shader->info.num_input_vgprs -= num_prolog_vgprs;
+
+ if (shader->key.as_ls || ctx->stage == MESA_SHADER_TESS_CTRL) {
+ if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
+ /* The LSHS size is not known until draw time, so we append it
+ * at the end of whatever LDS use there may be in the rest of
+ * the shader (currently none, unless LLVM decides to do its
+ * own LDS-based lowering).
+ */
+ ctx->ac.lds = LLVMAddGlobalInAddressSpace(ctx->ac.module, LLVMArrayType(ctx->ac.i32, 0),
+ "__lds_end", AC_ADDR_SPACE_LDS);
+ LLVMSetAlignment(ctx->ac.lds, 256);
+ } else {
+ ac_declare_lds_as_pointer(&ctx->ac);
+ }
+ }
+
+ /* Unlike radv, we override these arguments in the prolog, so to the
+ * API shader they appear as normal arguments.
+ */
+ if (ctx->stage == MESA_SHADER_VERTEX) {
+ ctx->abi.vertex_id = ac_get_arg(&ctx->ac, ctx->args.vertex_id);
+ ctx->abi.instance_id = ac_get_arg(&ctx->ac, ctx->args.instance_id);
+ } else if (ctx->stage == MESA_SHADER_FRAGMENT) {
+ ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args.persp_centroid);
+ ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args.linear_centroid);
+ }
}
-static void declare_const_and_shader_buffers(struct si_shader_context *ctx,
- bool assign_params)
-{
- enum ac_arg_type const_shader_buf_type;
+/* For the UMR disassembler. */
+#define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
+#define DEBUGGER_NUM_MARKERS 5
+
+static bool si_shader_binary_open(struct si_screen *screen, struct si_shader *shader,
+ struct ac_rtld_binary *rtld)
+{
+ const struct si_shader_selector *sel = shader->selector;
+ const char *part_elfs[5];
+ size_t part_sizes[5];
+ unsigned num_parts = 0;
+
+#define add_part(shader_or_part) \
+ if (shader_or_part) { \
+ part_elfs[num_parts] = (shader_or_part)->binary.elf_buffer; \
+ part_sizes[num_parts] = (shader_or_part)->binary.elf_size; \
+ num_parts++; \
+ }
+
+ add_part(shader->prolog);
+ add_part(shader->previous_stage);
+ add_part(shader->prolog2);
+ add_part(shader);
+ add_part(shader->epilog);
- if (ctx->shader->selector->info.const_buffers_declared == 1 &&
- ctx->shader->selector->info.shader_buffers_declared == 0)
- const_shader_buf_type = AC_ARG_CONST_FLOAT_PTR;
- else
- const_shader_buf_type = AC_ARG_CONST_DESC_PTR;
+#undef add_part
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, const_shader_buf_type,
- assign_params ? &ctx->const_and_shader_buffers :
- &ctx->other_const_and_shader_buffers);
+ struct ac_rtld_symbol lds_symbols[2];
+ unsigned num_lds_symbols = 0;
+
+ if (sel && screen->info.chip_class >= GFX9 && !shader->is_gs_copy_shader &&
+ (sel->info.stage == MESA_SHADER_GEOMETRY || shader->key.as_ngg)) {
+ /* We add this symbol even on LLVM <= 8 to ensure that
+ * shader->config.lds_size is set correctly below.
+ */
+ struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
+ sym->name = "esgs_ring";
+ sym->size = shader->gs_info.esgs_ring_size * 4;
+ sym->align = 64 * 1024;
+ }
+
+ if (shader->key.as_ngg && sel->info.stage == MESA_SHADER_GEOMETRY) {
+ struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
+ sym->name = "ngg_emit";
+ sym->size = shader->ngg.ngg_emit_size * 4;
+ sym->align = 4;
+ }
+
+ bool ok = ac_rtld_open(
+ rtld, (struct ac_rtld_open_info){.info = &screen->info,
+ .options =
+ {
+ .halt_at_entry = screen->options.halt_shaders,
+ },
+ .shader_type = sel->info.stage,
+ .wave_size = si_get_shader_wave_size(shader),
+ .num_parts = num_parts,
+ .elf_ptrs = part_elfs,
+ .elf_sizes = part_sizes,
+ .num_shared_lds_symbols = num_lds_symbols,
+ .shared_lds_symbols = lds_symbols});
+
+ if (rtld->lds_size > 0) {
+ unsigned alloc_granularity = screen->info.chip_class >= GFX7 ? 512 : 256;
+ shader->config.lds_size = align(rtld->lds_size, alloc_granularity) / alloc_granularity;
+ }
+
+ return ok;
}
-static void declare_samplers_and_images(struct si_shader_context *ctx,
- bool assign_params)
+static unsigned si_get_shader_binary_size(struct si_screen *screen, struct si_shader *shader)
{
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
- assign_params ? &ctx->samplers_and_images :
- &ctx->other_samplers_and_images);
+ struct ac_rtld_binary rtld;
+ si_shader_binary_open(screen, shader, &rtld);
+ return rtld.exec_size;
}
-static void declare_per_stage_desc_pointers(struct si_shader_context *ctx,
- bool assign_params)
+static bool si_get_external_symbol(void *data, const char *name, uint64_t *value)
{
- declare_const_and_shader_buffers(ctx, assign_params);
- declare_samplers_and_images(ctx, assign_params);
-}
+ uint64_t *scratch_va = data;
-static void declare_global_desc_pointers(struct si_shader_context *ctx)
-{
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR,
- &ctx->rw_buffers);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_IMAGE_PTR,
- &ctx->bindless_samplers_and_images);
-}
+ if (!strcmp(scratch_rsrc_dword0_symbol, name)) {
+ *value = (uint32_t)*scratch_va;
+ return true;
+ }
+ if (!strcmp(scratch_rsrc_dword1_symbol, name)) {
+ /* Enable scratch coalescing. */
+ *value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) | S_008F04_SWIZZLE_ENABLE(1);
+ return true;
+ }
-static void declare_vs_specific_input_sgprs(struct si_shader_context *ctx)
-{
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
- if (!ctx->shader->is_gs_copy_shader) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.base_vertex);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.start_instance);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.draw_id);
- }
+ return false;
}
-static void declare_vb_descriptor_input_sgprs(struct si_shader_context *ctx)
+bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
+ uint64_t scratch_va)
{
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_CONST_DESC_PTR, &ctx->vertex_buffers);
+ struct ac_rtld_binary binary;
+ if (!si_shader_binary_open(sscreen, shader, &binary))
+ return false;
- unsigned num_vbos_in_user_sgprs = ctx->shader->selector->num_vbos_in_user_sgprs;
- if (num_vbos_in_user_sgprs) {
- unsigned user_sgprs = ctx->args.num_sgprs_used;
+ si_resource_reference(&shader->bo, NULL);
+ shader->bo = si_aligned_buffer_create(
+ &sscreen->b, sscreen->info.cpdma_prefetch_writes_memory ? 0 : SI_RESOURCE_FLAG_READ_ONLY,
+ PIPE_USAGE_IMMUTABLE, align(binary.rx_size, SI_CPDMA_ALIGNMENT), 256);
+ if (!shader->bo)
+ return false;
- if (si_is_merged_shader(ctx))
- user_sgprs -= 8;
- assert(user_sgprs <= SI_SGPR_VS_VB_DESCRIPTOR_FIRST);
+ /* Upload. */
+ struct ac_rtld_upload_info u = {};
+ u.binary = &binary;
+ u.get_external_symbol = si_get_external_symbol;
+ u.cb_data = &scratch_va;
+ u.rx_va = shader->bo->gpu_address;
+ u.rx_ptr = sscreen->ws->buffer_map(
+ shader->bo->buf, NULL,
+ PIPE_TRANSFER_READ_WRITE | PIPE_TRANSFER_UNSYNCHRONIZED | RADEON_TRANSFER_TEMPORARY);
+ if (!u.rx_ptr)
+ return false;
- /* Declare unused SGPRs to align VB descriptors to 4 SGPRs (hw requirement). */
- for (unsigned i = user_sgprs; i < SI_SGPR_VS_VB_DESCRIPTOR_FIRST; i++)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
+ bool ok = ac_rtld_upload(&u);
- assert(num_vbos_in_user_sgprs <= ARRAY_SIZE(ctx->vb_descriptors));
- for (unsigned i = 0; i < num_vbos_in_user_sgprs; i++)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 4, AC_ARG_INT, &ctx->vb_descriptors[i]);
- }
-}
-
-static void declare_vs_input_vgprs(struct si_shader_context *ctx,
- unsigned *num_prolog_vgprs)
-{
- struct si_shader *shader = ctx->shader;
-
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.vertex_id);
- if (shader->key.as_ls) {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->rel_auto_id);
- if (ctx->screen->info.chip_class >= GFX10) {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
- } else {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
- }
- } else if (ctx->screen->info.chip_class >= GFX10) {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* user VGPR */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
- &ctx->vs_prim_id); /* user vgpr or PrimID (legacy) */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
- } else {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.instance_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->vs_prim_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL); /* unused */
- }
-
- if (!shader->is_gs_copy_shader) {
- /* Vertex load indices. */
- if (shader->selector->info.num_inputs) {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
- &ctx->vertex_index0);
- for (unsigned i = 1; i < shader->selector->info.num_inputs; i++)
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, NULL);
- }
- *num_prolog_vgprs += shader->selector->info.num_inputs;
- }
-}
-
-static void declare_vs_blit_inputs(struct si_shader_context *ctx,
- unsigned vs_blit_property)
-{
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
- &ctx->vs_blit_inputs); /* i16 x1, y1 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* i16 x1, y1 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* depth */
-
- if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_COLOR) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color0 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color1 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color2 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* color3 */
- } else if (vs_blit_property == SI_VS_BLIT_SGPRS_POS_TEXCOORD) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x1 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y1 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.x2 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.y2 */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.z */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL); /* texcoord.w */
- }
-}
+ sscreen->ws->buffer_unmap(shader->bo->buf);
+ ac_rtld_close(&binary);
-static void declare_tes_input_vgprs(struct si_shader_context *ctx)
-{
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_u);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, &ctx->tes_v);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->tes_rel_patch_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tes_patch_id);
+ return ok;
}
-enum {
- /* Convenient merged shader definitions. */
- SI_SHADER_MERGED_VERTEX_TESSCTRL = PIPE_SHADER_TYPES,
- SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY,
-};
+static void si_shader_dump_disassembly(struct si_screen *screen,
+ const struct si_shader_binary *binary,
+ gl_shader_stage stage, unsigned wave_size,
+ struct pipe_debug_callback *debug, const char *name,
+ FILE *file)
+{
+ struct ac_rtld_binary rtld_binary;
+
+ if (!ac_rtld_open(&rtld_binary, (struct ac_rtld_open_info){
+ .info = &screen->info,
+ .shader_type = stage,
+ .wave_size = wave_size,
+ .num_parts = 1,
+ .elf_ptrs = &binary->elf_buffer,
+ .elf_sizes = &binary->elf_size}))
+ return;
+
+ const char *disasm;
+ size_t nbytes;
+
+ if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm, &nbytes))
+ goto out;
+
+ if (nbytes > INT_MAX)
+ goto out;
+
+ if (debug && debug->debug_message) {
+ /* Very long debug messages are cut off, so send the
+ * disassembly one line at a time. This causes more
+ * overhead, but on the plus side it simplifies
+ * parsing of resulting logs.
+ */
+ pipe_debug_message(debug, SHADER_INFO, "Shader Disassembly Begin");
+
+ uint64_t line = 0;
+ while (line < nbytes) {
+ int count = nbytes - line;
+ const char *nl = memchr(disasm + line, '\n', nbytes - line);
+ if (nl)
+ count = nl - (disasm + line);
+
+ if (count) {
+ pipe_debug_message(debug, SHADER_INFO, "%.*s", count, disasm + line);
+ }
+
+ line += count + 1;
+ }
+
+ pipe_debug_message(debug, SHADER_INFO, "Shader Disassembly End");
+ }
+
+ if (file) {
+ fprintf(file, "Shader %s disassembly:\n", name);
+ fprintf(file, "%*s", (int)nbytes, disasm);
+ }
-void si_add_arg_checked(struct ac_shader_args *args,
- enum ac_arg_regfile file,
- unsigned registers, enum ac_arg_type type,
- struct ac_arg *arg,
- unsigned idx)
-{
- assert(args->arg_count == idx);
- ac_add_arg(args, file, registers, type, arg);
+out:
+ ac_rtld_close(&rtld_binary);
}
-void si_create_function(struct si_shader_context *ctx)
+static void si_calculate_max_simd_waves(struct si_shader *shader)
{
- struct si_shader *shader = ctx->shader;
- LLVMTypeRef returns[AC_MAX_ARGS];
- unsigned i, num_return_sgprs;
- unsigned num_returns = 0;
- unsigned num_prolog_vgprs = 0;
- unsigned type = ctx->type;
- unsigned vs_blit_property =
- shader->selector->info.properties[TGSI_PROPERTY_VS_BLIT_SGPRS_AMD];
-
- memset(&ctx->args, 0, sizeof(ctx->args));
-
- /* Set MERGED shaders. */
- if (ctx->screen->info.chip_class >= GFX9) {
- if (shader->key.as_ls || type == PIPE_SHADER_TESS_CTRL)
- type = SI_SHADER_MERGED_VERTEX_TESSCTRL; /* LS or HS */
- else if (shader->key.as_es || shader->key.as_ngg || type == PIPE_SHADER_GEOMETRY)
- type = SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY;
- }
-
- switch (type) {
- case PIPE_SHADER_VERTEX:
- declare_global_desc_pointers(ctx);
-
- if (vs_blit_property) {
- declare_vs_blit_inputs(ctx, vs_blit_property);
-
- /* VGPRs */
- declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
- break;
- }
-
- declare_per_stage_desc_pointers(ctx, true);
- declare_vs_specific_input_sgprs(ctx);
- if (!shader->is_gs_copy_shader)
- declare_vb_descriptor_input_sgprs(ctx);
-
- if (shader->key.as_es) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
- &ctx->es2gs_offset);
- } else if (shader->key.as_ls) {
- /* no extra parameters */
- } else {
- /* The locations of the other parameters are assigned dynamically. */
- declare_streamout_params(ctx, &shader->selector->so);
- }
-
- /* VGPRs */
- declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
-
- /* Return values */
- if (shader->key.opt.vs_as_prim_discard_cs) {
- for (i = 0; i < 4; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- }
- break;
-
- case PIPE_SHADER_TESS_CTRL: /* GFX6-GFX8 */
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx, true);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
-
- /* VGPRs */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
-
- /* param_tcs_offchip_offset and param_tcs_factor_offset are
- * placed after the user SGPRs.
- */
- for (i = 0; i < GFX6_TCS_NUM_USER_SGPR + 2; i++)
- returns[num_returns++] = ctx->i32; /* SGPRs */
- for (i = 0; i < 11; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- break;
-
- case SI_SHADER_MERGED_VERTEX_TESSCTRL:
- /* Merged stages have 8 system SGPRs at the beginning. */
- /* SPI_SHADER_USER_DATA_ADDR_LO/HI_HS */
- declare_per_stage_desc_pointers(ctx,
- ctx->type == PIPE_SHADER_TESS_CTRL);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_factor_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused */
-
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx,
- ctx->type == PIPE_SHADER_VERTEX);
- declare_vs_specific_input_sgprs(ctx);
-
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_offsets);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_out_lds_layout);
- declare_vb_descriptor_input_sgprs(ctx);
-
- /* VGPRs (first TCS, then VS) */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_patch_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.tcs_rel_ids);
-
- if (ctx->type == PIPE_SHADER_VERTEX) {
- declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
-
- /* LS return values are inputs to the TCS main shader part. */
- for (i = 0; i < 8 + GFX9_TCS_NUM_USER_SGPR; i++)
- returns[num_returns++] = ctx->i32; /* SGPRs */
- for (i = 0; i < 2; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- } else {
- /* TCS return values are inputs to the TCS epilog.
- *
- * param_tcs_offchip_offset, param_tcs_factor_offset,
- * param_tcs_offchip_layout, and param_rw_buffers
- * should be passed to the epilog.
- */
- for (i = 0; i <= 8 + GFX9_SGPR_TCS_OUT_LAYOUT; i++)
- returns[num_returns++] = ctx->i32; /* SGPRs */
- for (i = 0; i < 11; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- }
- break;
-
- case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
- /* Merged stages have 8 system SGPRs at the beginning. */
- /* SPI_SHADER_USER_DATA_ADDR_LO/HI_GS */
- declare_per_stage_desc_pointers(ctx,
- ctx->type == PIPE_SHADER_GEOMETRY);
-
- if (ctx->shader->key.as_ngg)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_tg_info);
- else
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
-
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_wave_info);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->merged_scratch_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS << 8) */
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL); /* unused (SPI_SHADER_PGM_LO/HI_GS >> 24) */
-
- declare_global_desc_pointers(ctx);
- if (ctx->type != PIPE_SHADER_VERTEX || !vs_blit_property) {
- declare_per_stage_desc_pointers(ctx,
- (ctx->type == PIPE_SHADER_VERTEX ||
- ctx->type == PIPE_SHADER_TESS_EVAL));
- }
-
- if (ctx->type == PIPE_SHADER_VERTEX) {
- if (vs_blit_property)
- declare_vs_blit_inputs(ctx, vs_blit_property);
- else
- declare_vs_specific_input_sgprs(ctx);
- } else {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
- /* Declare as many input SGPRs as the VS has. */
- }
-
- if (ctx->type == PIPE_SHADER_VERTEX)
- declare_vb_descriptor_input_sgprs(ctx);
-
- /* VGPRs (first GS, then VS/TES) */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx01_offset);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx23_offset);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx45_offset);
-
- if (ctx->type == PIPE_SHADER_VERTEX) {
- declare_vs_input_vgprs(ctx, &num_prolog_vgprs);
- } else if (ctx->type == PIPE_SHADER_TESS_EVAL) {
- declare_tes_input_vgprs(ctx);
- }
-
- if (ctx->shader->key.as_es &&
- (ctx->type == PIPE_SHADER_VERTEX ||
- ctx->type == PIPE_SHADER_TESS_EVAL)) {
- unsigned num_user_sgprs;
-
- if (ctx->type == PIPE_SHADER_VERTEX)
- num_user_sgprs = GFX9_VSGS_NUM_USER_SGPR;
- else
- num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
-
- /* ES return values are inputs to GS. */
- for (i = 0; i < 8 + num_user_sgprs; i++)
- returns[num_returns++] = ctx->i32; /* SGPRs */
- for (i = 0; i < 5; i++)
- returns[num_returns++] = ctx->f32; /* VGPRs */
- }
- break;
-
- case PIPE_SHADER_TESS_EVAL:
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx, true);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->vs_state_bits);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_layout);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tes_offchip_addr);
-
- if (shader->key.as_es) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->es2gs_offset);
- } else {
- declare_streamout_params(ctx, &shader->selector->so);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->tcs_offchip_offset);
- }
-
- /* VGPRs */
- declare_tes_input_vgprs(ctx);
- break;
-
- case PIPE_SHADER_GEOMETRY:
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx, true);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs2vs_offset);
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->gs_wave_id);
-
- /* VGPRs */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[0]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[1]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_prim_id);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[2]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[3]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[4]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->gs_vtx_offset[5]);
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &ctx->args.gs_invocation_id);
- break;
-
- case PIPE_SHADER_FRAGMENT:
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx, true);
- si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL,
- SI_PARAM_ALPHA_REF);
- si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
- &ctx->args.prim_mask, SI_PARAM_PRIM_MASK);
-
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT, &ctx->args.persp_sample,
- SI_PARAM_PERSP_SAMPLE);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
- &ctx->args.persp_center, SI_PARAM_PERSP_CENTER);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
- &ctx->args.persp_centroid, SI_PARAM_PERSP_CENTROID);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
- NULL, SI_PARAM_PERSP_PULL_MODEL);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
- &ctx->args.linear_sample, SI_PARAM_LINEAR_SAMPLE);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
- &ctx->args.linear_center, SI_PARAM_LINEAR_CENTER);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 2, AC_ARG_INT,
- &ctx->args.linear_centroid, SI_PARAM_LINEAR_CENTROID);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_FLOAT,
- NULL, SI_PARAM_LINE_STIPPLE_TEX);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
- &ctx->args.frag_pos[0], SI_PARAM_POS_X_FLOAT);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
- &ctx->args.frag_pos[1], SI_PARAM_POS_Y_FLOAT);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
- &ctx->args.frag_pos[2], SI_PARAM_POS_Z_FLOAT);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
- &ctx->args.frag_pos[3], SI_PARAM_POS_W_FLOAT);
- shader->info.face_vgpr_index = ctx->args.num_vgprs_used;
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
- &ctx->args.front_face, SI_PARAM_FRONT_FACE);
- shader->info.ancillary_vgpr_index = ctx->args.num_vgprs_used;
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
- &ctx->args.ancillary, SI_PARAM_ANCILLARY);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT,
- &ctx->args.sample_coverage, SI_PARAM_SAMPLE_COVERAGE);
- si_add_arg_checked(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT,
- &ctx->pos_fixed_pt, SI_PARAM_POS_FIXED_PT);
-
- /* Color inputs from the prolog. */
- if (shader->selector->info.colors_read) {
- unsigned num_color_elements =
- util_bitcount(shader->selector->info.colors_read);
-
- for (i = 0; i < num_color_elements; i++)
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
-
- num_prolog_vgprs += num_color_elements;
- }
-
- /* Outputs for the epilog. */
- num_return_sgprs = SI_SGPR_ALPHA_REF + 1;
- num_returns =
- num_return_sgprs +
- util_bitcount(shader->selector->info.colors_written) * 4 +
- shader->selector->info.writes_z +
- shader->selector->info.writes_stencil +
- shader->selector->info.writes_samplemask +
- 1 /* SampleMaskIn */;
-
- num_returns = MAX2(num_returns,
- num_return_sgprs +
- PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
-
- for (i = 0; i < num_return_sgprs; i++)
- returns[i] = ctx->i32;
- for (; i < num_returns; i++)
- returns[i] = ctx->f32;
- break;
-
- case PIPE_SHADER_COMPUTE:
- declare_global_desc_pointers(ctx);
- declare_per_stage_desc_pointers(ctx, true);
- if (shader->selector->info.uses_grid_size)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT,
- &ctx->args.num_work_groups);
- if (shader->selector->info.uses_block_size &&
- shader->selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 3, AC_ARG_INT, &ctx->block_size);
-
- unsigned cs_user_data_dwords =
- shader->selector->info.properties[TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD];
- if (cs_user_data_dwords) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, cs_user_data_dwords, AC_ARG_INT,
- &ctx->cs_user_data);
- }
-
- /* Hardware SGPRs. */
- for (i = 0; i < 3; i++) {
- if (shader->selector->info.uses_block_id[i]) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
- &ctx->args.workgroup_ids[i]);
- }
- }
- if (shader->selector->info.uses_subgroup_info)
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->args.tg_size);
-
- /* Hardware VGPRs. */
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 3, AC_ARG_INT,
- &ctx->args.local_invocation_ids);
- break;
- default:
- assert(0 && "unimplemented shader");
- return;
- }
-
- si_llvm_create_func(ctx, "main", returns, num_returns,
- si_get_max_workgroup_size(shader));
-
- /* Reserve register locations for VGPR inputs the PS prolog may need. */
- if (ctx->type == PIPE_SHADER_FRAGMENT && !ctx->shader->is_monolithic) {
- ac_llvm_add_target_dep_function_attr(ctx->main_fn,
- "InitialPSInputAddr",
- S_0286D0_PERSP_SAMPLE_ENA(1) |
- S_0286D0_PERSP_CENTER_ENA(1) |
- S_0286D0_PERSP_CENTROID_ENA(1) |
- S_0286D0_LINEAR_SAMPLE_ENA(1) |
- S_0286D0_LINEAR_CENTER_ENA(1) |
- S_0286D0_LINEAR_CENTROID_ENA(1) |
- S_0286D0_FRONT_FACE_ENA(1) |
- S_0286D0_ANCILLARY_ENA(1) |
- S_0286D0_POS_FIXED_PT_ENA(1));
- }
-
- shader->info.num_input_sgprs = ctx->args.num_sgprs_used;
- shader->info.num_input_vgprs = ctx->args.num_vgprs_used;
-
- assert(shader->info.num_input_vgprs >= num_prolog_vgprs);
- shader->info.num_input_vgprs -= num_prolog_vgprs;
-
- if (shader->key.as_ls || ctx->type == PIPE_SHADER_TESS_CTRL) {
- if (USE_LDS_SYMBOLS && LLVM_VERSION_MAJOR >= 9) {
- /* The LSHS size is not known until draw time, so we append it
- * at the end of whatever LDS use there may be in the rest of
- * the shader (currently none, unless LLVM decides to do its
- * own LDS-based lowering).
- */
- ctx->ac.lds = LLVMAddGlobalInAddressSpace(
- ctx->ac.module, LLVMArrayType(ctx->i32, 0),
- "__lds_end", AC_ADDR_SPACE_LDS);
- LLVMSetAlignment(ctx->ac.lds, 256);
- } else {
- ac_declare_lds_as_pointer(&ctx->ac);
- }
- }
-
- /* Unlike radv, we override these arguments in the prolog, so to the
- * API shader they appear as normal arguments.
- */
- if (ctx->type == PIPE_SHADER_VERTEX) {
- ctx->abi.vertex_id = ac_get_arg(&ctx->ac, ctx->args.vertex_id);
- ctx->abi.instance_id = ac_get_arg(&ctx->ac, ctx->args.instance_id);
- } else if (ctx->type == PIPE_SHADER_FRAGMENT) {
- ctx->abi.persp_centroid = ac_get_arg(&ctx->ac, ctx->args.persp_centroid);
- ctx->abi.linear_centroid = ac_get_arg(&ctx->ac, ctx->args.linear_centroid);
- }
+ struct si_screen *sscreen = shader->selector->screen;
+ struct ac_shader_config *conf = &shader->config;
+ unsigned num_inputs = shader->selector->info.num_inputs;
+ unsigned lds_increment = sscreen->info.chip_class >= GFX7 ? 512 : 256;
+ unsigned lds_per_wave = 0;
+ unsigned max_simd_waves;
+
+ max_simd_waves = sscreen->info.max_wave64_per_simd;
+
+ /* Compute LDS usage for PS. */
+ switch (shader->selector->info.stage) {
+ case MESA_SHADER_FRAGMENT:
+ /* The minimum usage per wave is (num_inputs * 48). The maximum
+ * usage is (num_inputs * 48 * 16).
+ * We can get anything in between and it varies between waves.
+ *
+ * The 48 bytes per input for a single primitive is equal to
+ * 4 bytes/component * 4 components/input * 3 points.
+ *
+ * Other stages don't know the size at compile time or don't
+ * allocate LDS per wave, but instead they do it per thread group.
+ */
+ lds_per_wave = conf->lds_size * lds_increment + align(num_inputs * 48, lds_increment);
+ break;
+ case MESA_SHADER_COMPUTE:
+ if (shader->selector) {
+ unsigned max_workgroup_size = si_get_max_workgroup_size(shader);
+ lds_per_wave = (conf->lds_size * lds_increment) /
+ DIV_ROUND_UP(max_workgroup_size, sscreen->compute_wave_size);
+ }
+ break;
+ default:;
+ }
+
+ /* Compute the per-SIMD wave counts. */
+ if (conf->num_sgprs) {
+ max_simd_waves =
+ MIN2(max_simd_waves, sscreen->info.num_physical_sgprs_per_simd / conf->num_sgprs);
+ }
+
+ if (conf->num_vgprs) {
+ /* Always print wave limits as Wave64, so that we can compare
+ * Wave32 and Wave64 with shader-db fairly. */
+ unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd;
+ max_simd_waves = MIN2(max_simd_waves, max_vgprs / conf->num_vgprs);
+ }
+
+ unsigned max_lds_per_simd = sscreen->info.lds_size_per_workgroup / 4;
+ if (lds_per_wave)
+ max_simd_waves = MIN2(max_simd_waves, max_lds_per_simd / lds_per_wave);
+
+ shader->info.max_simd_waves = max_simd_waves;
+}
+
+void si_shader_dump_stats_for_shader_db(struct si_screen *screen, struct si_shader *shader,
+ struct pipe_debug_callback *debug)
+{
+ const struct ac_shader_config *conf = &shader->config;
+
+ if (screen->options.debug_disassembly)
+ si_shader_dump_disassembly(screen, &shader->binary, shader->selector->info.stage,
+ si_get_shader_wave_size(shader), debug, "main", NULL);
+
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
+ "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
+ "Spilled VGPRs: %d PrivMem VGPRs: %d",
+ conf->num_sgprs, conf->num_vgprs, si_get_shader_binary_size(screen, shader),
+ conf->lds_size, conf->scratch_bytes_per_wave, shader->info.max_simd_waves,
+ conf->spilled_sgprs, conf->spilled_vgprs, shader->info.private_mem_vgprs);
+}
+
+static void si_shader_dump_stats(struct si_screen *sscreen, struct si_shader *shader, FILE *file,
+ bool check_debug_option)
+{
+ const struct ac_shader_config *conf = &shader->config;
+
+ if (!check_debug_option || si_can_dump_shader(sscreen, shader->selector->info.stage)) {
+ if (shader->selector->info.stage == MESA_SHADER_FRAGMENT) {
+ fprintf(file,
+ "*** SHADER CONFIG ***\n"
+ "SPI_PS_INPUT_ADDR = 0x%04x\n"
+ "SPI_PS_INPUT_ENA = 0x%04x\n",
+ conf->spi_ps_input_addr, conf->spi_ps_input_ena);
+ }
+
+ fprintf(file,
+ "*** SHADER STATS ***\n"
+ "SGPRS: %d\n"
+ "VGPRS: %d\n"
+ "Spilled SGPRs: %d\n"
+ "Spilled VGPRs: %d\n"
+ "Private memory VGPRs: %d\n"
+ "Code Size: %d bytes\n"
+ "LDS: %d blocks\n"
+ "Scratch: %d bytes per wave\n"
+ "Max Waves: %d\n"
+ "********************\n\n\n",
+ conf->num_sgprs, conf->num_vgprs, conf->spilled_sgprs, conf->spilled_vgprs,
+ shader->info.private_mem_vgprs, si_get_shader_binary_size(sscreen, shader),
+ conf->lds_size, conf->scratch_bytes_per_wave, shader->info.max_simd_waves);
+ }
}
-/* For the UMR disassembler. */
-#define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
-#define DEBUGGER_NUM_MARKERS 5
-
-static bool si_shader_binary_open(struct si_screen *screen,
- struct si_shader *shader,
- struct ac_rtld_binary *rtld)
+const char *si_get_shader_name(const struct si_shader *shader)
{
- const struct si_shader_selector *sel = shader->selector;
- const char *part_elfs[5];
- size_t part_sizes[5];
- unsigned num_parts = 0;
-
-#define add_part(shader_or_part) \
- if (shader_or_part) { \
- part_elfs[num_parts] = (shader_or_part)->binary.elf_buffer; \
- part_sizes[num_parts] = (shader_or_part)->binary.elf_size; \
- num_parts++; \
- }
-
- add_part(shader->prolog);
- add_part(shader->previous_stage);
- add_part(shader->prolog2);
- add_part(shader);
- add_part(shader->epilog);
-
-#undef add_part
-
- struct ac_rtld_symbol lds_symbols[2];
- unsigned num_lds_symbols = 0;
-
- if (sel && screen->info.chip_class >= GFX9 && !shader->is_gs_copy_shader &&
- (sel->type == PIPE_SHADER_GEOMETRY || shader->key.as_ngg)) {
- /* We add this symbol even on LLVM <= 8 to ensure that
- * shader->config.lds_size is set correctly below.
- */
- struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
- sym->name = "esgs_ring";
- sym->size = shader->gs_info.esgs_ring_size;
- sym->align = 64 * 1024;
- }
-
- if (shader->key.as_ngg && sel->type == PIPE_SHADER_GEOMETRY) {
- struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
- sym->name = "ngg_emit";
- sym->size = shader->ngg.ngg_emit_size * 4;
- sym->align = 4;
- }
-
- bool ok = ac_rtld_open(rtld, (struct ac_rtld_open_info){
- .info = &screen->info,
- .options = {
- .halt_at_entry = screen->options.halt_shaders,
- },
- .shader_type = tgsi_processor_to_shader_stage(sel->type),
- .wave_size = si_get_shader_wave_size(shader),
- .num_parts = num_parts,
- .elf_ptrs = part_elfs,
- .elf_sizes = part_sizes,
- .num_shared_lds_symbols = num_lds_symbols,
- .shared_lds_symbols = lds_symbols });
-
- if (rtld->lds_size > 0) {
- unsigned alloc_granularity = screen->info.chip_class >= GFX7 ? 512 : 256;
- shader->config.lds_size =
- align(rtld->lds_size, alloc_granularity) / alloc_granularity;
- }
-
- return ok;
+ switch (shader->selector->info.stage) {
+ case MESA_SHADER_VERTEX:
+ if (shader->key.as_es)
+ return "Vertex Shader as ES";
+ else if (shader->key.as_ls)
+ return "Vertex Shader as LS";
+ else if (shader->key.opt.vs_as_prim_discard_cs)
+ return "Vertex Shader as Primitive Discard CS";
+ else if (shader->key.as_ngg)
+ return "Vertex Shader as ESGS";
+ else
+ return "Vertex Shader as VS";
+ case MESA_SHADER_TESS_CTRL:
+ return "Tessellation Control Shader";
+ case MESA_SHADER_TESS_EVAL:
+ if (shader->key.as_es)
+ return "Tessellation Evaluation Shader as ES";
+ else if (shader->key.as_ngg)
+ return "Tessellation Evaluation Shader as ESGS";
+ else
+ return "Tessellation Evaluation Shader as VS";
+ case MESA_SHADER_GEOMETRY:
+ if (shader->is_gs_copy_shader)
+ return "GS Copy Shader as VS";
+ else
+ return "Geometry Shader";
+ case MESA_SHADER_FRAGMENT:
+ return "Pixel Shader";
+ case MESA_SHADER_COMPUTE:
+ return "Compute Shader";
+ default:
+ return "Unknown Shader";
+ }
}
-static unsigned si_get_shader_binary_size(struct si_screen *screen, struct si_shader *shader)
+void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
+ struct pipe_debug_callback *debug, FILE *file, bool check_debug_option)
{
- struct ac_rtld_binary rtld;
- si_shader_binary_open(screen, shader, &rtld);
- return rtld.exec_size;
-}
+ gl_shader_stage stage = shader->selector->info.stage;
-static bool si_get_external_symbol(void *data, const char *name, uint64_t *value)
-{
- uint64_t *scratch_va = data;
-
- if (!strcmp(scratch_rsrc_dword0_symbol, name)) {
- *value = (uint32_t)*scratch_va;
- return true;
- }
- if (!strcmp(scratch_rsrc_dword1_symbol, name)) {
- /* Enable scratch coalescing. */
- *value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) |
- S_008F04_SWIZZLE_ENABLE(1);
- return true;
- }
-
- return false;
-}
+ if (!check_debug_option || si_can_dump_shader(sscreen, stage))
+ si_dump_shader_key(shader, file);
-bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader *shader,
- uint64_t scratch_va)
-{
- struct ac_rtld_binary binary;
- if (!si_shader_binary_open(sscreen, shader, &binary))
- return false;
-
- si_resource_reference(&shader->bo, NULL);
- shader->bo = si_aligned_buffer_create(&sscreen->b,
- sscreen->info.cpdma_prefetch_writes_memory ?
- 0 : SI_RESOURCE_FLAG_READ_ONLY,
- PIPE_USAGE_IMMUTABLE,
- align(binary.rx_size, SI_CPDMA_ALIGNMENT),
- 256);
- if (!shader->bo)
- return false;
-
- /* Upload. */
- struct ac_rtld_upload_info u = {};
- u.binary = &binary;
- u.get_external_symbol = si_get_external_symbol;
- u.cb_data = &scratch_va;
- u.rx_va = shader->bo->gpu_address;
- u.rx_ptr = sscreen->ws->buffer_map(shader->bo->buf, NULL,
- PIPE_TRANSFER_READ_WRITE |
- PIPE_TRANSFER_UNSYNCHRONIZED |
- RADEON_TRANSFER_TEMPORARY);
- if (!u.rx_ptr)
- return false;
-
- bool ok = ac_rtld_upload(&u);
-
- sscreen->ws->buffer_unmap(shader->bo->buf);
- ac_rtld_close(&binary);
-
- return ok;
-}
+ if (!check_debug_option && shader->binary.llvm_ir_string) {
+ if (shader->previous_stage && shader->previous_stage->binary.llvm_ir_string) {
+ fprintf(file, "\n%s - previous stage - LLVM IR:\n\n", si_get_shader_name(shader));
+ fprintf(file, "%s\n", shader->previous_stage->binary.llvm_ir_string);
+ }
-static void si_shader_dump_disassembly(struct si_screen *screen,
- const struct si_shader_binary *binary,
- enum pipe_shader_type shader_type,
- unsigned wave_size,
- struct pipe_debug_callback *debug,
- const char *name, FILE *file)
-{
- struct ac_rtld_binary rtld_binary;
-
- if (!ac_rtld_open(&rtld_binary, (struct ac_rtld_open_info){
- .info = &screen->info,
- .shader_type = tgsi_processor_to_shader_stage(shader_type),
- .wave_size = wave_size,
- .num_parts = 1,
- .elf_ptrs = &binary->elf_buffer,
- .elf_sizes = &binary->elf_size }))
- return;
-
- const char *disasm;
- size_t nbytes;
-
- if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm, &nbytes))
- goto out;
-
- if (nbytes > INT_MAX)
- goto out;
-
- if (debug && debug->debug_message) {
- /* Very long debug messages are cut off, so send the
- * disassembly one line at a time. This causes more
- * overhead, but on the plus side it simplifies
- * parsing of resulting logs.
- */
- pipe_debug_message(debug, SHADER_INFO,
- "Shader Disassembly Begin");
-
- uint64_t line = 0;
- while (line < nbytes) {
- int count = nbytes - line;
- const char *nl = memchr(disasm + line, '\n', nbytes - line);
- if (nl)
- count = nl - (disasm + line);
-
- if (count) {
- pipe_debug_message(debug, SHADER_INFO,
- "%.*s", count, disasm + line);
- }
-
- line += count + 1;
- }
-
- pipe_debug_message(debug, SHADER_INFO,
- "Shader Disassembly End");
- }
-
- if (file) {
- fprintf(file, "Shader %s disassembly:\n", name);
- fprintf(file, "%*s", (int)nbytes, disasm);
- }
+ fprintf(file, "\n%s - main shader part - LLVM IR:\n\n", si_get_shader_name(shader));
+ fprintf(file, "%s\n", shader->binary.llvm_ir_string);
+ }
-out:
- ac_rtld_close(&rtld_binary);
-}
+ if (!check_debug_option ||
+ (si_can_dump_shader(sscreen, stage) && !(sscreen->debug_flags & DBG(NO_ASM)))) {
+ unsigned wave_size = si_get_shader_wave_size(shader);
-static void si_calculate_max_simd_waves(struct si_shader *shader)
-{
- struct si_screen *sscreen = shader->selector->screen;
- struct ac_shader_config *conf = &shader->config;
- unsigned num_inputs = shader->selector->info.num_inputs;
- unsigned lds_increment = sscreen->info.chip_class >= GFX7 ? 512 : 256;
- unsigned lds_per_wave = 0;
- unsigned max_simd_waves;
-
- max_simd_waves = sscreen->info.max_wave64_per_simd;
-
- /* Compute LDS usage for PS. */
- switch (shader->selector->type) {
- case PIPE_SHADER_FRAGMENT:
- /* The minimum usage per wave is (num_inputs * 48). The maximum
- * usage is (num_inputs * 48 * 16).
- * We can get anything in between and it varies between waves.
- *
- * The 48 bytes per input for a single primitive is equal to
- * 4 bytes/component * 4 components/input * 3 points.
- *
- * Other stages don't know the size at compile time or don't
- * allocate LDS per wave, but instead they do it per thread group.
- */
- lds_per_wave = conf->lds_size * lds_increment +
- align(num_inputs * 48, lds_increment);
- break;
- case PIPE_SHADER_COMPUTE:
- if (shader->selector) {
- unsigned max_workgroup_size =
- si_get_max_workgroup_size(shader);
- lds_per_wave = (conf->lds_size * lds_increment) /
- DIV_ROUND_UP(max_workgroup_size,
- sscreen->compute_wave_size);
- }
- break;
- default:;
- }
-
- /* Compute the per-SIMD wave counts. */
- if (conf->num_sgprs) {
- max_simd_waves =
- MIN2(max_simd_waves,
- sscreen->info.num_physical_sgprs_per_simd / conf->num_sgprs);
- }
-
- if (conf->num_vgprs) {
- /* Always print wave limits as Wave64, so that we can compare
- * Wave32 and Wave64 with shader-db fairly. */
- unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd;
- max_simd_waves = MIN2(max_simd_waves, max_vgprs / conf->num_vgprs);
- }
-
- /* LDS is 64KB per CU (4 SIMDs) on GFX6-9, which is 16KB per SIMD (usage above
- * 16KB makes some SIMDs unoccupied).
- *
- * LDS is 128KB in WGP mode and 64KB in CU mode. Assume the WGP mode is used.
- */
- unsigned max_lds_size = sscreen->info.chip_class >= GFX10 ? 128*1024 : 64*1024;
- unsigned max_lds_per_simd = max_lds_size / 4;
- if (lds_per_wave)
- max_simd_waves = MIN2(max_simd_waves, max_lds_per_simd / lds_per_wave);
-
- shader->info.max_simd_waves = max_simd_waves;
-}
+ fprintf(file, "\n%s:\n", si_get_shader_name(shader));
-void si_shader_dump_stats_for_shader_db(struct si_screen *screen,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
-{
- const struct ac_shader_config *conf = &shader->config;
-
- if (screen->options.debug_disassembly)
- si_shader_dump_disassembly(screen, &shader->binary,
- shader->selector->type,
- si_get_shader_wave_size(shader),
- debug, "main", NULL);
-
- pipe_debug_message(debug, SHADER_INFO,
- "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d "
- "LDS: %d Scratch: %d Max Waves: %d Spilled SGPRs: %d "
- "Spilled VGPRs: %d PrivMem VGPRs: %d",
- conf->num_sgprs, conf->num_vgprs,
- si_get_shader_binary_size(screen, shader),
- conf->lds_size, conf->scratch_bytes_per_wave,
- shader->info.max_simd_waves, conf->spilled_sgprs,
- conf->spilled_vgprs, shader->info.private_mem_vgprs);
-}
+ if (shader->prolog)
+ si_shader_dump_disassembly(sscreen, &shader->prolog->binary, stage, wave_size, debug,
+ "prolog", file);
+ if (shader->previous_stage)
+ si_shader_dump_disassembly(sscreen, &shader->previous_stage->binary, stage,
+ wave_size, debug, "previous stage", file);
+ if (shader->prolog2)
+ si_shader_dump_disassembly(sscreen, &shader->prolog2->binary, stage, wave_size,
+ debug, "prolog2", file);
-static void si_shader_dump_stats(struct si_screen *sscreen,
- struct si_shader *shader,
- FILE *file,
- bool check_debug_option)
-{
- const struct ac_shader_config *conf = &shader->config;
-
- if (!check_debug_option ||
- si_can_dump_shader(sscreen, shader->selector->type)) {
- if (shader->selector->type == PIPE_SHADER_FRAGMENT) {
- fprintf(file, "*** SHADER CONFIG ***\n"
- "SPI_PS_INPUT_ADDR = 0x%04x\n"
- "SPI_PS_INPUT_ENA = 0x%04x\n",
- conf->spi_ps_input_addr, conf->spi_ps_input_ena);
- }
-
- fprintf(file, "*** SHADER STATS ***\n"
- "SGPRS: %d\n"
- "VGPRS: %d\n"
- "Spilled SGPRs: %d\n"
- "Spilled VGPRs: %d\n"
- "Private memory VGPRs: %d\n"
- "Code Size: %d bytes\n"
- "LDS: %d blocks\n"
- "Scratch: %d bytes per wave\n"
- "Max Waves: %d\n"
- "********************\n\n\n",
- conf->num_sgprs, conf->num_vgprs,
- conf->spilled_sgprs, conf->spilled_vgprs,
- shader->info.private_mem_vgprs,
- si_get_shader_binary_size(sscreen, shader),
- conf->lds_size, conf->scratch_bytes_per_wave,
- shader->info.max_simd_waves);
- }
-}
+ si_shader_dump_disassembly(sscreen, &shader->binary, stage, wave_size, debug, "main",
+ file);
-const char *si_get_shader_name(const struct si_shader *shader)
-{
- switch (shader->selector->type) {
- case PIPE_SHADER_VERTEX:
- if (shader->key.as_es)
- return "Vertex Shader as ES";
- else if (shader->key.as_ls)
- return "Vertex Shader as LS";
- else if (shader->key.opt.vs_as_prim_discard_cs)
- return "Vertex Shader as Primitive Discard CS";
- else if (shader->key.as_ngg)
- return "Vertex Shader as ESGS";
- else
- return "Vertex Shader as VS";
- case PIPE_SHADER_TESS_CTRL:
- return "Tessellation Control Shader";
- case PIPE_SHADER_TESS_EVAL:
- if (shader->key.as_es)
- return "Tessellation Evaluation Shader as ES";
- else if (shader->key.as_ngg)
- return "Tessellation Evaluation Shader as ESGS";
- else
- return "Tessellation Evaluation Shader as VS";
- case PIPE_SHADER_GEOMETRY:
- if (shader->is_gs_copy_shader)
- return "GS Copy Shader as VS";
- else
- return "Geometry Shader";
- case PIPE_SHADER_FRAGMENT:
- return "Pixel Shader";
- case PIPE_SHADER_COMPUTE:
- return "Compute Shader";
- default:
- return "Unknown Shader";
- }
-}
+ if (shader->epilog)
+ si_shader_dump_disassembly(sscreen, &shader->epilog->binary, stage, wave_size, debug,
+ "epilog", file);
+ fprintf(file, "\n");
+ }
-void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
- struct pipe_debug_callback *debug,
- FILE *file, bool check_debug_option)
-{
- enum pipe_shader_type shader_type = shader->selector->type;
-
- if (!check_debug_option ||
- si_can_dump_shader(sscreen, shader_type))
- si_dump_shader_key(shader, file);
-
- if (!check_debug_option && shader->binary.llvm_ir_string) {
- if (shader->previous_stage &&
- shader->previous_stage->binary.llvm_ir_string) {
- fprintf(file, "\n%s - previous stage - LLVM IR:\n\n",
- si_get_shader_name(shader));
- fprintf(file, "%s\n", shader->previous_stage->binary.llvm_ir_string);
- }
-
- fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
- si_get_shader_name(shader));
- fprintf(file, "%s\n", shader->binary.llvm_ir_string);
- }
-
- if (!check_debug_option ||
- (si_can_dump_shader(sscreen, shader_type) &&
- !(sscreen->debug_flags & DBG(NO_ASM)))) {
- unsigned wave_size = si_get_shader_wave_size(shader);
-
- fprintf(file, "\n%s:\n", si_get_shader_name(shader));
-
- if (shader->prolog)
- si_shader_dump_disassembly(sscreen, &shader->prolog->binary,
- shader_type, wave_size, debug, "prolog", file);
- if (shader->previous_stage)
- si_shader_dump_disassembly(sscreen, &shader->previous_stage->binary,
- shader_type, wave_size, debug, "previous stage", file);
- if (shader->prolog2)
- si_shader_dump_disassembly(sscreen, &shader->prolog2->binary,
- shader_type, wave_size, debug, "prolog2", file);
-
- si_shader_dump_disassembly(sscreen, &shader->binary, shader_type,
- wave_size, debug, "main", file);
-
- if (shader->epilog)
- si_shader_dump_disassembly(sscreen, &shader->epilog->binary,
- shader_type, wave_size, debug, "epilog", file);
- fprintf(file, "\n");
- }
-
- si_shader_dump_stats(sscreen, shader, file, check_debug_option);
+ si_shader_dump_stats(sscreen, shader, file, check_debug_option);
}
static void si_dump_shader_key_vs(const struct si_shader_key *key,
- const struct si_vs_prolog_bits *prolog,
- const char *prefix, FILE *f)
-{
- fprintf(f, " %s.instance_divisor_is_one = %u\n",
- prefix, prolog->instance_divisor_is_one);
- fprintf(f, " %s.instance_divisor_is_fetched = %u\n",
- prefix, prolog->instance_divisor_is_fetched);
- fprintf(f, " %s.unpack_instance_id_from_vertex_id = %u\n",
- prefix, prolog->unpack_instance_id_from_vertex_id);
- fprintf(f, " %s.ls_vgpr_fix = %u\n",
- prefix, prolog->ls_vgpr_fix);
-
- fprintf(f, " mono.vs.fetch_opencode = %x\n", key->mono.vs_fetch_opencode);
- fprintf(f, " mono.vs.fix_fetch = {");
- for (int i = 0; i < SI_MAX_ATTRIBS; i++) {
- union si_vs_fix_fetch fix = key->mono.vs_fix_fetch[i];
- if (i)
- fprintf(f, ", ");
- if (!fix.bits)
- fprintf(f, "0");
- else
- fprintf(f, "%u.%u.%u.%u", fix.u.reverse, fix.u.log_size,
- fix.u.num_channels_m1, fix.u.format);
- }
- fprintf(f, "}\n");
+ const struct si_vs_prolog_bits *prolog, const char *prefix,
+ FILE *f)
+{
+ fprintf(f, " %s.instance_divisor_is_one = %u\n", prefix, prolog->instance_divisor_is_one);
+ fprintf(f, " %s.instance_divisor_is_fetched = %u\n", prefix,
+ prolog->instance_divisor_is_fetched);
+ fprintf(f, " %s.unpack_instance_id_from_vertex_id = %u\n", prefix,
+ prolog->unpack_instance_id_from_vertex_id);
+ fprintf(f, " %s.ls_vgpr_fix = %u\n", prefix, prolog->ls_vgpr_fix);
+
+ fprintf(f, " mono.vs.fetch_opencode = %x\n", key->mono.vs_fetch_opencode);
+ fprintf(f, " mono.vs.fix_fetch = {");
+ for (int i = 0; i < SI_MAX_ATTRIBS; i++) {
+ union si_vs_fix_fetch fix = key->mono.vs_fix_fetch[i];
+ if (i)
+ fprintf(f, ", ");
+ if (!fix.bits)
+ fprintf(f, "0");
+ else
+ fprintf(f, "%u.%u.%u.%u", fix.u.reverse, fix.u.log_size, fix.u.num_channels_m1,
+ fix.u.format);
+ }
+ fprintf(f, "}\n");
}
static void si_dump_shader_key(const struct si_shader *shader, FILE *f)
{
- const struct si_shader_key *key = &shader->key;
- enum pipe_shader_type shader_type = shader->selector->type;
-
- fprintf(f, "SHADER KEY\n");
-
- switch (shader_type) {
- case PIPE_SHADER_VERTEX:
- si_dump_shader_key_vs(key, &key->part.vs.prolog,
- "part.vs.prolog", f);
- fprintf(f, " as_es = %u\n", key->as_es);
- fprintf(f, " as_ls = %u\n", key->as_ls);
- fprintf(f, " as_ngg = %u\n", key->as_ngg);
- fprintf(f, " mono.u.vs_export_prim_id = %u\n",
- key->mono.u.vs_export_prim_id);
- fprintf(f, " opt.vs_as_prim_discard_cs = %u\n",
- key->opt.vs_as_prim_discard_cs);
- fprintf(f, " opt.cs_prim_type = %s\n",
- tgsi_primitive_names[key->opt.cs_prim_type]);
- fprintf(f, " opt.cs_indexed = %u\n",
- key->opt.cs_indexed);
- fprintf(f, " opt.cs_instancing = %u\n",
- key->opt.cs_instancing);
- fprintf(f, " opt.cs_primitive_restart = %u\n",
- key->opt.cs_primitive_restart);
- fprintf(f, " opt.cs_provoking_vertex_first = %u\n",
- key->opt.cs_provoking_vertex_first);
- fprintf(f, " opt.cs_need_correct_orientation = %u\n",
- key->opt.cs_need_correct_orientation);
- fprintf(f, " opt.cs_cull_front = %u\n",
- key->opt.cs_cull_front);
- fprintf(f, " opt.cs_cull_back = %u\n",
- key->opt.cs_cull_back);
- fprintf(f, " opt.cs_cull_z = %u\n",
- key->opt.cs_cull_z);
- fprintf(f, " opt.cs_halfz_clip_space = %u\n",
- key->opt.cs_halfz_clip_space);
- break;
-
- case PIPE_SHADER_TESS_CTRL:
- if (shader->selector->screen->info.chip_class >= GFX9) {
- si_dump_shader_key_vs(key, &key->part.tcs.ls_prolog,
- "part.tcs.ls_prolog", f);
- }
- fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
- fprintf(f, " mono.u.ff_tcs_inputs_to_copy = 0x%"PRIx64"\n", key->mono.u.ff_tcs_inputs_to_copy);
- break;
-
- case PIPE_SHADER_TESS_EVAL:
- fprintf(f, " as_es = %u\n", key->as_es);
- fprintf(f, " as_ngg = %u\n", key->as_ngg);
- fprintf(f, " mono.u.vs_export_prim_id = %u\n",
- key->mono.u.vs_export_prim_id);
- break;
-
- case PIPE_SHADER_GEOMETRY:
- if (shader->is_gs_copy_shader)
- break;
-
- if (shader->selector->screen->info.chip_class >= GFX9 &&
- key->part.gs.es->type == PIPE_SHADER_VERTEX) {
- si_dump_shader_key_vs(key, &key->part.gs.vs_prolog,
- "part.gs.vs_prolog", f);
- }
- fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n", key->part.gs.prolog.tri_strip_adj_fix);
- fprintf(f, " part.gs.prolog.gfx9_prev_is_vs = %u\n", key->part.gs.prolog.gfx9_prev_is_vs);
- fprintf(f, " as_ngg = %u\n", key->as_ngg);
- break;
-
- case PIPE_SHADER_COMPUTE:
- break;
-
- case PIPE_SHADER_FRAGMENT:
- fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
- fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
- fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
- fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n", key->part.ps.prolog.force_persp_sample_interp);
- fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n", key->part.ps.prolog.force_linear_sample_interp);
- fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n", key->part.ps.prolog.force_persp_center_interp);
- fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n", key->part.ps.prolog.force_linear_center_interp);
- fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n", key->part.ps.prolog.bc_optimize_for_persp);
- fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n", key->part.ps.prolog.bc_optimize_for_linear);
- fprintf(f, " part.ps.prolog.samplemask_log_ps_iter = %u\n", key->part.ps.prolog.samplemask_log_ps_iter);
- fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n", key->part.ps.epilog.spi_shader_col_format);
- fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
- fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
- fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
- fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
- fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
- fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n", key->part.ps.epilog.poly_line_smoothing);
- fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
- fprintf(f, " mono.u.ps.interpolate_at_sample_force_center = %u\n", key->mono.u.ps.interpolate_at_sample_force_center);
- fprintf(f, " mono.u.ps.fbfetch_msaa = %u\n", key->mono.u.ps.fbfetch_msaa);
- fprintf(f, " mono.u.ps.fbfetch_is_1D = %u\n", key->mono.u.ps.fbfetch_is_1D);
- fprintf(f, " mono.u.ps.fbfetch_layered = %u\n", key->mono.u.ps.fbfetch_layered);
- break;
-
- default:
- assert(0);
- }
-
- if ((shader_type == PIPE_SHADER_GEOMETRY ||
- shader_type == PIPE_SHADER_TESS_EVAL ||
- shader_type == PIPE_SHADER_VERTEX) &&
- !key->as_es && !key->as_ls) {
- fprintf(f, " opt.kill_outputs = 0x%"PRIx64"\n", key->opt.kill_outputs);
- fprintf(f, " opt.clip_disable = %u\n", key->opt.clip_disable);
- }
+ const struct si_shader_key *key = &shader->key;
+ gl_shader_stage stage = shader->selector->info.stage;
+
+ fprintf(f, "SHADER KEY\n");
+
+ switch (stage) {
+ case MESA_SHADER_VERTEX:
+ si_dump_shader_key_vs(key, &key->part.vs.prolog, "part.vs.prolog", f);
+ fprintf(f, " as_es = %u\n", key->as_es);
+ fprintf(f, " as_ls = %u\n", key->as_ls);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
+ fprintf(f, " mono.u.vs_export_prim_id = %u\n", key->mono.u.vs_export_prim_id);
+ fprintf(f, " opt.vs_as_prim_discard_cs = %u\n", key->opt.vs_as_prim_discard_cs);
+ fprintf(f, " opt.cs_prim_type = %s\n", tgsi_primitive_names[key->opt.cs_prim_type]);
+ fprintf(f, " opt.cs_indexed = %u\n", key->opt.cs_indexed);
+ fprintf(f, " opt.cs_instancing = %u\n", key->opt.cs_instancing);
+ fprintf(f, " opt.cs_primitive_restart = %u\n", key->opt.cs_primitive_restart);
+ fprintf(f, " opt.cs_provoking_vertex_first = %u\n", key->opt.cs_provoking_vertex_first);
+ fprintf(f, " opt.cs_need_correct_orientation = %u\n", key->opt.cs_need_correct_orientation);
+ fprintf(f, " opt.cs_cull_front = %u\n", key->opt.cs_cull_front);
+ fprintf(f, " opt.cs_cull_back = %u\n", key->opt.cs_cull_back);
+ fprintf(f, " opt.cs_cull_z = %u\n", key->opt.cs_cull_z);
+ fprintf(f, " opt.cs_halfz_clip_space = %u\n", key->opt.cs_halfz_clip_space);
+ break;
+
+ case MESA_SHADER_TESS_CTRL:
+ if (shader->selector->screen->info.chip_class >= GFX9) {
+ si_dump_shader_key_vs(key, &key->part.tcs.ls_prolog, "part.tcs.ls_prolog", f);
+ }
+ fprintf(f, " part.tcs.epilog.prim_mode = %u\n", key->part.tcs.epilog.prim_mode);
+ fprintf(f, " mono.u.ff_tcs_inputs_to_copy = 0x%" PRIx64 "\n",
+ key->mono.u.ff_tcs_inputs_to_copy);
+ break;
+
+ case MESA_SHADER_TESS_EVAL:
+ fprintf(f, " as_es = %u\n", key->as_es);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
+ fprintf(f, " mono.u.vs_export_prim_id = %u\n", key->mono.u.vs_export_prim_id);
+ break;
+
+ case MESA_SHADER_GEOMETRY:
+ if (shader->is_gs_copy_shader)
+ break;
+
+ if (shader->selector->screen->info.chip_class >= GFX9 &&
+ key->part.gs.es->info.stage == MESA_SHADER_VERTEX) {
+ si_dump_shader_key_vs(key, &key->part.gs.vs_prolog, "part.gs.vs_prolog", f);
+ }
+ fprintf(f, " part.gs.prolog.tri_strip_adj_fix = %u\n",
+ key->part.gs.prolog.tri_strip_adj_fix);
+ fprintf(f, " part.gs.prolog.gfx9_prev_is_vs = %u\n", key->part.gs.prolog.gfx9_prev_is_vs);
+ fprintf(f, " as_ngg = %u\n", key->as_ngg);
+ break;
+
+ case MESA_SHADER_COMPUTE:
+ break;
+
+ case MESA_SHADER_FRAGMENT:
+ fprintf(f, " part.ps.prolog.color_two_side = %u\n", key->part.ps.prolog.color_two_side);
+ fprintf(f, " part.ps.prolog.flatshade_colors = %u\n", key->part.ps.prolog.flatshade_colors);
+ fprintf(f, " part.ps.prolog.poly_stipple = %u\n", key->part.ps.prolog.poly_stipple);
+ fprintf(f, " part.ps.prolog.force_persp_sample_interp = %u\n",
+ key->part.ps.prolog.force_persp_sample_interp);
+ fprintf(f, " part.ps.prolog.force_linear_sample_interp = %u\n",
+ key->part.ps.prolog.force_linear_sample_interp);
+ fprintf(f, " part.ps.prolog.force_persp_center_interp = %u\n",
+ key->part.ps.prolog.force_persp_center_interp);
+ fprintf(f, " part.ps.prolog.force_linear_center_interp = %u\n",
+ key->part.ps.prolog.force_linear_center_interp);
+ fprintf(f, " part.ps.prolog.bc_optimize_for_persp = %u\n",
+ key->part.ps.prolog.bc_optimize_for_persp);
+ fprintf(f, " part.ps.prolog.bc_optimize_for_linear = %u\n",
+ key->part.ps.prolog.bc_optimize_for_linear);
+ fprintf(f, " part.ps.prolog.samplemask_log_ps_iter = %u\n",
+ key->part.ps.prolog.samplemask_log_ps_iter);
+ fprintf(f, " part.ps.epilog.spi_shader_col_format = 0x%x\n",
+ key->part.ps.epilog.spi_shader_col_format);
+ fprintf(f, " part.ps.epilog.color_is_int8 = 0x%X\n", key->part.ps.epilog.color_is_int8);
+ fprintf(f, " part.ps.epilog.color_is_int10 = 0x%X\n", key->part.ps.epilog.color_is_int10);
+ fprintf(f, " part.ps.epilog.last_cbuf = %u\n", key->part.ps.epilog.last_cbuf);
+ fprintf(f, " part.ps.epilog.alpha_func = %u\n", key->part.ps.epilog.alpha_func);
+ fprintf(f, " part.ps.epilog.alpha_to_one = %u\n", key->part.ps.epilog.alpha_to_one);
+ fprintf(f, " part.ps.epilog.poly_line_smoothing = %u\n",
+ key->part.ps.epilog.poly_line_smoothing);
+ fprintf(f, " part.ps.epilog.clamp_color = %u\n", key->part.ps.epilog.clamp_color);
+ fprintf(f, " mono.u.ps.interpolate_at_sample_force_center = %u\n",
+ key->mono.u.ps.interpolate_at_sample_force_center);
+ fprintf(f, " mono.u.ps.fbfetch_msaa = %u\n", key->mono.u.ps.fbfetch_msaa);
+ fprintf(f, " mono.u.ps.fbfetch_is_1D = %u\n", key->mono.u.ps.fbfetch_is_1D);
+ fprintf(f, " mono.u.ps.fbfetch_layered = %u\n", key->mono.u.ps.fbfetch_layered);
+ break;
+
+ default:
+ assert(0);
+ }
+
+ if ((stage == MESA_SHADER_GEOMETRY || stage == MESA_SHADER_TESS_EVAL ||
+ stage == MESA_SHADER_VERTEX) &&
+ !key->as_es && !key->as_ls) {
+ fprintf(f, " opt.kill_outputs = 0x%" PRIx64 "\n", key->opt.kill_outputs);
+ fprintf(f, " opt.clip_disable = %u\n", key->opt.clip_disable);
+ if (stage != MESA_SHADER_GEOMETRY)
+ fprintf(f, " opt.ngg_culling = 0x%x\n", key->opt.ngg_culling);
+ }
}
static void si_optimize_vs_outputs(struct si_shader_context *ctx)
{
- struct si_shader *shader = ctx->shader;
- struct si_shader_info *info = &shader->selector->info;
-
- if ((ctx->type != PIPE_SHADER_VERTEX &&
- ctx->type != PIPE_SHADER_TESS_EVAL) ||
- shader->key.as_ls ||
- shader->key.as_es)
- return;
-
- ac_optimize_vs_outputs(&ctx->ac,
- ctx->main_fn,
- shader->info.vs_output_param_offset,
- info->num_outputs,
- &shader->info.nr_param_exports);
-}
+ struct si_shader *shader = ctx->shader;
+ struct si_shader_info *info = &shader->selector->info;
+ unsigned skip_vs_optim_mask = 0;
-static void si_init_exec_from_input(struct si_shader_context *ctx,
- struct ac_arg param, unsigned bitoffset)
-{
- LLVMValueRef args[] = {
- ac_get_arg(&ctx->ac, param),
- LLVMConstInt(ctx->i32, bitoffset, 0),
- };
- ac_build_intrinsic(&ctx->ac,
- "llvm.amdgcn.init.exec.from.input",
- ctx->voidt, args, 2, AC_FUNC_ATTR_CONVERGENT);
-}
+ if ((ctx->stage != MESA_SHADER_VERTEX && ctx->stage != MESA_SHADER_TESS_EVAL) ||
+ shader->key.as_ls || shader->key.as_es)
+ return;
-static bool si_vs_needs_prolog(const struct si_shader_selector *sel,
- const struct si_vs_prolog_bits *key)
-{
- /* VGPR initialization fixup for Vega10 and Raven is always done in the
- * VS prolog. */
- return sel->vs_needs_prolog ||
- key->ls_vgpr_fix ||
- key->unpack_instance_id_from_vertex_id;
+ /* Optimizing these outputs is not possible, since they might be overriden
+ * at runtime with S_028644_PT_SPRITE_TEX. */
+ for (int i = 0; i < info->num_outputs; i++) {
+ if (info->output_semantic_name[i] == TGSI_SEMANTIC_PCOORD ||
+ info->output_semantic_name[i] == TGSI_SEMANTIC_TEXCOORD) {
+ skip_vs_optim_mask |= 1u << shader->info.vs_output_param_offset[i];
+ }
+ }
+
+ ac_optimize_vs_outputs(&ctx->ac, ctx->main_fn, shader->info.vs_output_param_offset,
+ info->num_outputs, skip_vs_optim_mask,
+ &shader->info.nr_param_exports);
}
-static bool si_build_main_function(struct si_shader_context *ctx,
- struct nir_shader *nir, bool free_nir)
-{
- struct si_shader *shader = ctx->shader;
- struct si_shader_selector *sel = shader->selector;
-
- si_llvm_init_resource_callbacks(ctx);
-
- switch (ctx->type) {
- case PIPE_SHADER_VERTEX:
- if (shader->key.as_ls)
- ctx->abi.emit_outputs = si_llvm_emit_ls_epilogue;
- else if (shader->key.as_es)
- ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
- else if (shader->key.opt.vs_as_prim_discard_cs)
- ctx->abi.emit_outputs = si_llvm_emit_prim_discard_cs_epilogue;
- else if (shader->key.as_ngg)
- ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue;
- else
- ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
- ctx->abi.load_base_vertex = get_base_vertex;
- break;
- case PIPE_SHADER_TESS_CTRL:
- si_llvm_init_tcs_callbacks(ctx);
- break;
- case PIPE_SHADER_TESS_EVAL:
- si_llvm_init_tes_callbacks(ctx);
-
- if (shader->key.as_es)
- ctx->abi.emit_outputs = si_llvm_emit_es_epilogue;
- else if (shader->key.as_ngg)
- ctx->abi.emit_outputs = gfx10_emit_ngg_epilogue;
- else
- ctx->abi.emit_outputs = si_llvm_emit_vs_epilogue;
- break;
- case PIPE_SHADER_GEOMETRY:
- si_llvm_init_gs_callbacks(ctx);
- break;
- case PIPE_SHADER_FRAGMENT:
- si_llvm_init_ps_callbacks(ctx);
- break;
- case PIPE_SHADER_COMPUTE:
- ctx->abi.load_local_group_size = get_block_size;
- break;
- default:
- assert(!"Unsupported shader type");
- return false;
- }
-
- si_create_function(ctx);
-
- if (ctx->shader->key.as_es || ctx->type == PIPE_SHADER_GEOMETRY)
- si_preload_esgs_ring(ctx);
-
- if (ctx->type == PIPE_SHADER_GEOMETRY)
- si_preload_gs_rings(ctx);
- else if (ctx->type == PIPE_SHADER_TESS_EVAL)
- si_llvm_preload_tes_rings(ctx);
-
- if (ctx->type == PIPE_SHADER_TESS_CTRL &&
- sel->info.tessfactors_are_def_in_all_invocs) {
- for (unsigned i = 0; i < 6; i++) {
- ctx->invoc0_tess_factors[i] =
- ac_build_alloca_undef(&ctx->ac, ctx->i32, "");
- }
- }
-
- if (ctx->type == PIPE_SHADER_GEOMETRY) {
- for (unsigned i = 0; i < 4; i++) {
- ctx->gs_next_vertex[i] =
- ac_build_alloca(&ctx->ac, ctx->i32, "");
- }
- if (shader->key.as_ngg) {
- for (unsigned i = 0; i < 4; ++i) {
- ctx->gs_curprim_verts[i] =
- ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
- ctx->gs_generated_prims[i] =
- ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
- }
-
- unsigned scratch_size = 8;
- if (sel->so.num_outputs)
- scratch_size = 44;
-
- LLVMTypeRef ai32 = LLVMArrayType(ctx->i32, scratch_size);
- ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
- ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
- LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(ai32));
- LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
-
- ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(ctx->ac.module,
- LLVMArrayType(ctx->i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
- LLVMSetLinkage(ctx->gs_ngg_emit, LLVMExternalLinkage);
- LLVMSetAlignment(ctx->gs_ngg_emit, 4);
- }
- }
-
- if (ctx->type != PIPE_SHADER_GEOMETRY &&
- (shader->key.as_ngg && !shader->key.as_es)) {
- /* Unconditionally declare scratch space base for streamout and
- * vertex compaction. Whether space is actually allocated is
- * determined during linking / PM4 creation.
- *
- * Add an extra dword per vertex to ensure an odd stride, which
- * avoids bank conflicts for SoA accesses.
- */
- if (!gfx10_is_ngg_passthrough(shader))
- si_llvm_declare_esgs_ring(ctx);
-
- /* This is really only needed when streamout and / or vertex
- * compaction is enabled.
- */
- if (sel->so.num_outputs && !ctx->gs_ngg_scratch) {
- LLVMTypeRef asi32 = LLVMArrayType(ctx->i32, 8);
- ctx->gs_ngg_scratch = LLVMAddGlobalInAddressSpace(ctx->ac.module,
- asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
- LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32));
- LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
- }
- }
-
- /* For GFX9 merged shaders:
- * - Set EXEC for the first shader. If the prolog is present, set
- * EXEC there instead.
- * - Add a barrier before the second shader.
- * - In the second shader, reset EXEC to ~0 and wrap the main part in
- * an if-statement. This is required for correctness in geometry
- * shaders, to ensure that empty GS waves do not send GS_EMIT and
- * GS_CUT messages.
- *
- * For monolithic merged shaders, the first shader is wrapped in an
- * if-block together with its prolog in si_build_wrapper_function.
- *
- * NGG vertex and tess eval shaders running as the last
- * vertex/geometry stage handle execution explicitly using
- * if-statements.
- */
- if (ctx->screen->info.chip_class >= GFX9) {
- if (!shader->is_monolithic &&
- (shader->key.as_es || shader->key.as_ls) &&
- (ctx->type == PIPE_SHADER_TESS_EVAL ||
- (ctx->type == PIPE_SHADER_VERTEX &&
- !si_vs_needs_prolog(sel, &shader->key.part.vs.prolog)))) {
- si_init_exec_from_input(ctx,
- ctx->merged_wave_info, 0);
- } else if (ctx->type == PIPE_SHADER_TESS_CTRL ||
- ctx->type == PIPE_SHADER_GEOMETRY ||
- (shader->key.as_ngg && !shader->key.as_es)) {
- LLVMValueRef thread_enabled;
- bool nested_barrier;
-
- if (!shader->is_monolithic ||
- (ctx->type == PIPE_SHADER_TESS_EVAL &&
- (shader->key.as_ngg && !shader->key.as_es)))
- ac_init_exec_full_mask(&ctx->ac);
-
- if ((ctx->type == PIPE_SHADER_VERTEX ||
- ctx->type == PIPE_SHADER_TESS_EVAL) &&
- shader->key.as_ngg && !shader->key.as_es) {
- gfx10_ngg_build_sendmsg_gs_alloc_req(ctx);
-
- /* Build the primitive export at the beginning
- * of the shader if possible.
- */
- if (gfx10_ngg_export_prim_early(shader))
- gfx10_ngg_build_export_prim(ctx, NULL);
- }
-
- if (ctx->type == PIPE_SHADER_TESS_CTRL ||
- ctx->type == PIPE_SHADER_GEOMETRY) {
- if (ctx->type == PIPE_SHADER_GEOMETRY && shader->key.as_ngg) {
- gfx10_ngg_gs_emit_prologue(ctx);
- nested_barrier = false;
- } else {
- nested_barrier = true;
- }
-
- thread_enabled = si_is_gs_thread(ctx);
- } else {
- thread_enabled = si_is_es_thread(ctx);
- nested_barrier = false;
- }
-
- ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
- ctx->merged_wrap_if_label = 11500;
- ac_build_ifcc(&ctx->ac, thread_enabled, ctx->merged_wrap_if_label);
-
- if (nested_barrier) {
- /* Execute a barrier before the second shader in
- * a merged shader.
- *
- * Execute the barrier inside the conditional block,
- * so that empty waves can jump directly to s_endpgm,
- * which will also signal the barrier.
- *
- * This is possible in gfx9, because an empty wave
- * for the second shader does not participate in
- * the epilogue. With NGG, empty waves may still
- * be required to export data (e.g. GS output vertices),
- * so we cannot let them exit early.
- *
- * If the shader is TCS and the TCS epilog is present
- * and contains a barrier, it will wait there and then
- * reach s_endpgm.
- */
- si_llvm_emit_barrier(ctx);
- }
- }
- }
-
- if (sel->force_correct_derivs_after_kill) {
- ctx->postponed_kill = ac_build_alloca_undef(&ctx->ac, ctx->i1, "");
- /* true = don't kill. */
- LLVMBuildStore(ctx->ac.builder, ctx->i1true,
- ctx->postponed_kill);
- }
-
- bool success = si_nir_build_llvm(ctx, nir);
- if (free_nir)
- ralloc_free(nir);
- if (!success) {
- fprintf(stderr, "Failed to translate shader from NIR to LLVM\n");
- return false;
- }
-
- si_llvm_build_ret(ctx, ctx->return_value);
- return true;
+static bool si_vs_needs_prolog(const struct si_shader_selector *sel,
+ const struct si_vs_prolog_bits *prolog_key,
+ const struct si_shader_key *key, bool ngg_cull_shader)
+{
+ /* VGPR initialization fixup for Vega10 and Raven is always done in the
+ * VS prolog. */
+ return sel->vs_needs_prolog || prolog_key->ls_vgpr_fix ||
+ prolog_key->unpack_instance_id_from_vertex_id ||
+ (ngg_cull_shader && key->opt.ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL);
+}
+
+static bool si_build_main_function(struct si_shader_context *ctx, struct si_shader *shader,
+ struct nir_shader *nir, bool free_nir, bool ngg_cull_shader)
+{
+ struct si_shader_selector *sel = shader->selector;
+ const struct si_shader_info *info = &sel->info;
+
+ ctx->shader = shader;
+ ctx->stage = sel->info.stage;
+
+ ctx->num_const_buffers = util_last_bit(info->const_buffers_declared);
+ ctx->num_shader_buffers = util_last_bit(info->shader_buffers_declared);
+
+ ctx->num_samplers = util_last_bit(info->samplers_declared);
+ ctx->num_images = util_last_bit(info->images_declared);
+
+ si_llvm_init_resource_callbacks(ctx);
+
+ switch (ctx->stage) {
+ case MESA_SHADER_VERTEX:
+ si_llvm_init_vs_callbacks(ctx, ngg_cull_shader);
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ si_llvm_init_tcs_callbacks(ctx);
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ si_llvm_init_tes_callbacks(ctx, ngg_cull_shader);
+ break;
+ case MESA_SHADER_GEOMETRY:
+ si_llvm_init_gs_callbacks(ctx);
+ break;
+ case MESA_SHADER_FRAGMENT:
+ si_llvm_init_ps_callbacks(ctx);
+ break;
+ case MESA_SHADER_COMPUTE:
+ ctx->abi.load_local_group_size = si_llvm_get_block_size;
+ break;
+ default:
+ assert(!"Unsupported shader type");
+ return false;
+ }
+
+ si_create_function(ctx, ngg_cull_shader);
+
+ if (ctx->shader->key.as_es || ctx->stage == MESA_SHADER_GEOMETRY)
+ si_preload_esgs_ring(ctx);
+
+ if (ctx->stage == MESA_SHADER_GEOMETRY)
+ si_preload_gs_rings(ctx);
+ else if (ctx->stage == MESA_SHADER_TESS_EVAL)
+ si_llvm_preload_tes_rings(ctx);
+
+ if (ctx->stage == MESA_SHADER_TESS_CTRL && sel->info.tessfactors_are_def_in_all_invocs) {
+ for (unsigned i = 0; i < 6; i++) {
+ ctx->invoc0_tess_factors[i] = ac_build_alloca_undef(&ctx->ac, ctx->ac.i32, "");
+ }
+ }
+
+ if (ctx->stage == MESA_SHADER_GEOMETRY) {
+ for (unsigned i = 0; i < 4; i++) {
+ ctx->gs_next_vertex[i] = ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
+ }
+ if (shader->key.as_ngg) {
+ for (unsigned i = 0; i < 4; ++i) {
+ ctx->gs_curprim_verts[i] = ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
+ ctx->gs_generated_prims[i] = ac_build_alloca(&ctx->ac, ctx->ac.i32, "");
+ }
+
+ assert(!ctx->gs_ngg_scratch);
+ LLVMTypeRef ai32 = LLVMArrayType(ctx->ac.i32, gfx10_ngg_get_scratch_dw_size(shader));
+ ctx->gs_ngg_scratch =
+ LLVMAddGlobalInAddressSpace(ctx->ac.module, ai32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+ LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(ai32));
+ LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
+
+ ctx->gs_ngg_emit = LLVMAddGlobalInAddressSpace(
+ ctx->ac.module, LLVMArrayType(ctx->ac.i32, 0), "ngg_emit", AC_ADDR_SPACE_LDS);
+ LLVMSetLinkage(ctx->gs_ngg_emit, LLVMExternalLinkage);
+ LLVMSetAlignment(ctx->gs_ngg_emit, 4);
+ }
+ }
+
+ if (ctx->stage != MESA_SHADER_GEOMETRY && (shader->key.as_ngg && !shader->key.as_es)) {
+ /* Unconditionally declare scratch space base for streamout and
+ * vertex compaction. Whether space is actually allocated is
+ * determined during linking / PM4 creation.
+ *
+ * Add an extra dword per vertex to ensure an odd stride, which
+ * avoids bank conflicts for SoA accesses.
+ */
+ if (!gfx10_is_ngg_passthrough(shader))
+ si_llvm_declare_esgs_ring(ctx);
+
+ /* This is really only needed when streamout and / or vertex
+ * compaction is enabled.
+ */
+ if (!ctx->gs_ngg_scratch && (sel->so.num_outputs || shader->key.opt.ngg_culling)) {
+ LLVMTypeRef asi32 = LLVMArrayType(ctx->ac.i32, gfx10_ngg_get_scratch_dw_size(shader));
+ ctx->gs_ngg_scratch =
+ LLVMAddGlobalInAddressSpace(ctx->ac.module, asi32, "ngg_scratch", AC_ADDR_SPACE_LDS);
+ LLVMSetInitializer(ctx->gs_ngg_scratch, LLVMGetUndef(asi32));
+ LLVMSetAlignment(ctx->gs_ngg_scratch, 4);
+ }
+ }
+
+ /* For GFX9 merged shaders:
+ * - Set EXEC for the first shader. If the prolog is present, set
+ * EXEC there instead.
+ * - Add a barrier before the second shader.
+ * - In the second shader, reset EXEC to ~0 and wrap the main part in
+ * an if-statement. This is required for correctness in geometry
+ * shaders, to ensure that empty GS waves do not send GS_EMIT and
+ * GS_CUT messages.
+ *
+ * For monolithic merged shaders, the first shader is wrapped in an
+ * if-block together with its prolog in si_build_wrapper_function.
+ *
+ * NGG vertex and tess eval shaders running as the last
+ * vertex/geometry stage handle execution explicitly using
+ * if-statements.
+ */
+ if (ctx->screen->info.chip_class >= GFX9) {
+ if (!shader->is_monolithic && (shader->key.as_es || shader->key.as_ls) &&
+ (ctx->stage == MESA_SHADER_TESS_EVAL ||
+ (ctx->stage == MESA_SHADER_VERTEX &&
+ !si_vs_needs_prolog(sel, &shader->key.part.vs.prolog, &shader->key, ngg_cull_shader)))) {
+ si_init_exec_from_input(ctx, ctx->merged_wave_info, 0);
+ } else if (ctx->stage == MESA_SHADER_TESS_CTRL || ctx->stage == MESA_SHADER_GEOMETRY ||
+ (shader->key.as_ngg && !shader->key.as_es)) {
+ LLVMValueRef thread_enabled;
+ bool nested_barrier;
+
+ if (!shader->is_monolithic || (ctx->stage == MESA_SHADER_TESS_EVAL && shader->key.as_ngg &&
+ !shader->key.as_es && !shader->key.opt.ngg_culling))
+ ac_init_exec_full_mask(&ctx->ac);
+
+ if ((ctx->stage == MESA_SHADER_VERTEX || ctx->stage == MESA_SHADER_TESS_EVAL) &&
+ shader->key.as_ngg && !shader->key.as_es && !shader->key.opt.ngg_culling) {
+ gfx10_ngg_build_sendmsg_gs_alloc_req(ctx);
+
+ /* Build the primitive export at the beginning
+ * of the shader if possible.
+ */
+ if (gfx10_ngg_export_prim_early(shader))
+ gfx10_ngg_build_export_prim(ctx, NULL, NULL);
+ }
+
+ if (ctx->stage == MESA_SHADER_TESS_CTRL || ctx->stage == MESA_SHADER_GEOMETRY) {
+ if (ctx->stage == MESA_SHADER_GEOMETRY && shader->key.as_ngg) {
+ gfx10_ngg_gs_emit_prologue(ctx);
+ nested_barrier = false;
+ } else {
+ nested_barrier = true;
+ }
+
+ thread_enabled = si_is_gs_thread(ctx);
+ } else {
+ thread_enabled = si_is_es_thread(ctx);
+ nested_barrier = false;
+ }
+
+ ctx->merged_wrap_if_entry_block = LLVMGetInsertBlock(ctx->ac.builder);
+ ctx->merged_wrap_if_label = 11500;
+ ac_build_ifcc(&ctx->ac, thread_enabled, ctx->merged_wrap_if_label);
+
+ if (nested_barrier) {
+ /* Execute a barrier before the second shader in
+ * a merged shader.
+ *
+ * Execute the barrier inside the conditional block,
+ * so that empty waves can jump directly to s_endpgm,
+ * which will also signal the barrier.
+ *
+ * This is possible in gfx9, because an empty wave
+ * for the second shader does not participate in
+ * the epilogue. With NGG, empty waves may still
+ * be required to export data (e.g. GS output vertices),
+ * so we cannot let them exit early.
+ *
+ * If the shader is TCS and the TCS epilog is present
+ * and contains a barrier, it will wait there and then
+ * reach s_endpgm.
+ */
+ si_llvm_emit_barrier(ctx);
+ }
+ }
+ }
+
+ bool success = si_nir_build_llvm(ctx, nir);
+ if (free_nir)
+ ralloc_free(nir);
+ if (!success) {
+ fprintf(stderr, "Failed to translate shader from NIR to LLVM\n");
+ return false;
+ }
+
+ si_llvm_build_ret(ctx, ctx->return_value);
+ return true;
}
/**
*
* \param info Shader info of the vertex shader.
* \param num_input_sgprs Number of input SGPRs for the vertex shader.
+ * \param has_old_ Whether the preceding shader part is the NGG cull shader.
* \param prolog_key Key of the VS prolog
* \param shader_out The vertex shader, or the next shader if merging LS+HS or ES+GS.
* \param key Output shader part key.
*/
-static void si_get_vs_prolog_key(const struct si_shader_info *info,
- unsigned num_input_sgprs,
- const struct si_vs_prolog_bits *prolog_key,
- struct si_shader *shader_out,
- union si_shader_part_key *key)
-{
- memset(key, 0, sizeof(*key));
- key->vs_prolog.states = *prolog_key;
- key->vs_prolog.num_input_sgprs = num_input_sgprs;
- key->vs_prolog.num_inputs = info->num_inputs;
- key->vs_prolog.as_ls = shader_out->key.as_ls;
- key->vs_prolog.as_es = shader_out->key.as_es;
- key->vs_prolog.as_ngg = shader_out->key.as_ngg;
-
- if (shader_out->selector->type == PIPE_SHADER_TESS_CTRL) {
- key->vs_prolog.as_ls = 1;
- key->vs_prolog.num_merged_next_stage_vgprs = 2;
- } else if (shader_out->selector->type == PIPE_SHADER_GEOMETRY) {
- key->vs_prolog.as_es = 1;
- key->vs_prolog.num_merged_next_stage_vgprs = 5;
- } else if (shader_out->key.as_ngg) {
- key->vs_prolog.num_merged_next_stage_vgprs = 5;
- }
-
- /* Enable loading the InstanceID VGPR. */
- uint16_t input_mask = u_bit_consecutive(0, info->num_inputs);
-
- if ((key->vs_prolog.states.instance_divisor_is_one |
- key->vs_prolog.states.instance_divisor_is_fetched) & input_mask)
- shader_out->info.uses_instanceid = true;
-}
-
-/**
- * Given a list of shader part functions, build a wrapper function that
- * runs them in sequence to form a monolithic shader.
- */
-void si_build_wrapper_function(struct si_shader_context *ctx, LLVMValueRef *parts,
- unsigned num_parts, unsigned main_part,
- unsigned next_shader_first_part)
-{
- LLVMBuilderRef builder = ctx->ac.builder;
- /* PS epilog has one arg per color component; gfx9 merged shader
- * prologs need to forward 40 SGPRs.
- */
- LLVMValueRef initial[AC_MAX_ARGS], out[AC_MAX_ARGS];
- LLVMTypeRef function_type;
- unsigned num_first_params;
- unsigned num_out, initial_num_out;
- ASSERTED unsigned num_out_sgpr; /* used in debug checks */
- ASSERTED unsigned initial_num_out_sgpr; /* used in debug checks */
- unsigned num_sgprs, num_vgprs;
- unsigned gprs;
-
- memset(&ctx->args, 0, sizeof(ctx->args));
-
- for (unsigned i = 0; i < num_parts; ++i) {
- ac_add_function_attr(ctx->ac.context, parts[i], -1,
- AC_FUNC_ATTR_ALWAYSINLINE);
- LLVMSetLinkage(parts[i], LLVMPrivateLinkage);
- }
-
- /* The parameters of the wrapper function correspond to those of the
- * first part in terms of SGPRs and VGPRs, but we use the types of the
- * main part to get the right types. This is relevant for the
- * dereferenceable attribute on descriptor table pointers.
- */
- num_sgprs = 0;
- num_vgprs = 0;
-
- function_type = LLVMGetElementType(LLVMTypeOf(parts[0]));
- num_first_params = LLVMCountParamTypes(function_type);
-
- for (unsigned i = 0; i < num_first_params; ++i) {
- LLVMValueRef param = LLVMGetParam(parts[0], i);
-
- if (ac_is_sgpr_param(param)) {
- assert(num_vgprs == 0);
- num_sgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
- } else {
- num_vgprs += ac_get_type_size(LLVMTypeOf(param)) / 4;
- }
- }
-
- gprs = 0;
- while (gprs < num_sgprs + num_vgprs) {
- LLVMValueRef param = LLVMGetParam(parts[main_part], ctx->args.arg_count);
- LLVMTypeRef type = LLVMTypeOf(param);
- unsigned size = ac_get_type_size(type) / 4;
-
- /* This is going to get casted anyways, so we don't have to
- * have the exact same type. But we do have to preserve the
- * pointer-ness so that LLVM knows about it.
- */
- enum ac_arg_type arg_type = AC_ARG_INT;
- if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
- type = LLVMGetElementType(type);
-
- if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
- if (LLVMGetVectorSize(type) == 4)
- arg_type = AC_ARG_CONST_DESC_PTR;
- else if (LLVMGetVectorSize(type) == 8)
- arg_type = AC_ARG_CONST_IMAGE_PTR;
- else
- assert(0);
- } else if (type == ctx->f32) {
- arg_type = AC_ARG_CONST_FLOAT_PTR;
- } else {
- assert(0);
- }
- }
-
- ac_add_arg(&ctx->args, gprs < num_sgprs ? AC_ARG_SGPR : AC_ARG_VGPR,
- size, arg_type, NULL);
-
- assert(ac_is_sgpr_param(param) == (gprs < num_sgprs));
- assert(gprs + size <= num_sgprs + num_vgprs &&
- (gprs >= num_sgprs || gprs + size <= num_sgprs));
-
- gprs += size;
- }
-
- /* Prepare the return type. */
- unsigned num_returns = 0;
- LLVMTypeRef returns[AC_MAX_ARGS], last_func_type, return_type;
-
- last_func_type = LLVMGetElementType(LLVMTypeOf(parts[num_parts - 1]));
- return_type = LLVMGetReturnType(last_func_type);
-
- switch (LLVMGetTypeKind(return_type)) {
- case LLVMStructTypeKind:
- num_returns = LLVMCountStructElementTypes(return_type);
- assert(num_returns <= ARRAY_SIZE(returns));
- LLVMGetStructElementTypes(return_type, returns);
- break;
- case LLVMVoidTypeKind:
- break;
- default:
- unreachable("unexpected type");
- }
-
- si_llvm_create_func(ctx, "wrapper", returns, num_returns,
- si_get_max_workgroup_size(ctx->shader));
-
- if (si_is_merged_shader(ctx))
- ac_init_exec_full_mask(&ctx->ac);
-
- /* Record the arguments of the function as if they were an output of
- * a previous part.
- */
- num_out = 0;
- num_out_sgpr = 0;
-
- for (unsigned i = 0; i < ctx->args.arg_count; ++i) {
- LLVMValueRef param = LLVMGetParam(ctx->main_fn, i);
- LLVMTypeRef param_type = LLVMTypeOf(param);
- LLVMTypeRef out_type = ctx->args.args[i].file == AC_ARG_SGPR ? ctx->i32 : ctx->f32;
- unsigned size = ac_get_type_size(param_type) / 4;
-
- if (size == 1) {
- if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
- param = LLVMBuildPtrToInt(builder, param, ctx->i32, "");
- param_type = ctx->i32;
- }
-
- if (param_type != out_type)
- param = LLVMBuildBitCast(builder, param, out_type, "");
- out[num_out++] = param;
- } else {
- LLVMTypeRef vector_type = LLVMVectorType(out_type, size);
-
- if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
- param = LLVMBuildPtrToInt(builder, param, ctx->i64, "");
- param_type = ctx->i64;
- }
-
- if (param_type != vector_type)
- param = LLVMBuildBitCast(builder, param, vector_type, "");
-
- for (unsigned j = 0; j < size; ++j)
- out[num_out++] = LLVMBuildExtractElement(
- builder, param, LLVMConstInt(ctx->i32, j, 0), "");
- }
-
- if (ctx->args.args[i].file == AC_ARG_SGPR)
- num_out_sgpr = num_out;
- }
-
- memcpy(initial, out, sizeof(out));
- initial_num_out = num_out;
- initial_num_out_sgpr = num_out_sgpr;
-
- /* Now chain the parts. */
- LLVMValueRef ret = NULL;
- for (unsigned part = 0; part < num_parts; ++part) {
- LLVMValueRef in[AC_MAX_ARGS];
- LLVMTypeRef ret_type;
- unsigned out_idx = 0;
- unsigned num_params = LLVMCountParams(parts[part]);
-
- /* Merged shaders are executed conditionally depending
- * on the number of enabled threads passed in the input SGPRs. */
- if (is_multi_part_shader(ctx) && part == 0) {
- LLVMValueRef ena, count = initial[3];
-
- count = LLVMBuildAnd(builder, count,
- LLVMConstInt(ctx->i32, 0x7f, 0), "");
- ena = LLVMBuildICmp(builder, LLVMIntULT,
- ac_get_thread_id(&ctx->ac), count, "");
- ac_build_ifcc(&ctx->ac, ena, 6506);
- }
-
- /* Derive arguments for the next part from outputs of the
- * previous one.
- */
- for (unsigned param_idx = 0; param_idx < num_params; ++param_idx) {
- LLVMValueRef param;
- LLVMTypeRef param_type;
- bool is_sgpr;
- unsigned param_size;
- LLVMValueRef arg = NULL;
-
- param = LLVMGetParam(parts[part], param_idx);
- param_type = LLVMTypeOf(param);
- param_size = ac_get_type_size(param_type) / 4;
- is_sgpr = ac_is_sgpr_param(param);
-
- if (is_sgpr) {
- ac_add_function_attr(ctx->ac.context, parts[part],
- param_idx + 1, AC_FUNC_ATTR_INREG);
- } else if (out_idx < num_out_sgpr) {
- /* Skip returned SGPRs the current part doesn't
- * declare on the input. */
- out_idx = num_out_sgpr;
- }
-
- assert(out_idx + param_size <= (is_sgpr ? num_out_sgpr : num_out));
-
- if (param_size == 1)
- arg = out[out_idx];
- else
- arg = ac_build_gather_values(&ctx->ac, &out[out_idx], param_size);
-
- if (LLVMTypeOf(arg) != param_type) {
- if (LLVMGetTypeKind(param_type) == LLVMPointerTypeKind) {
- if (LLVMGetPointerAddressSpace(param_type) ==
- AC_ADDR_SPACE_CONST_32BIT) {
- arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
- arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
- } else {
- arg = LLVMBuildBitCast(builder, arg, ctx->i64, "");
- arg = LLVMBuildIntToPtr(builder, arg, param_type, "");
- }
- } else {
- arg = LLVMBuildBitCast(builder, arg, param_type, "");
- }
- }
-
- in[param_idx] = arg;
- out_idx += param_size;
- }
-
- ret = ac_build_call(&ctx->ac, parts[part], in, num_params);
-
- if (is_multi_part_shader(ctx) &&
- part + 1 == next_shader_first_part) {
- ac_build_endif(&ctx->ac, 6506);
-
- /* The second half of the merged shader should use
- * the inputs from the toplevel (wrapper) function,
- * not the return value from the last call.
- *
- * That's because the last call was executed condi-
- * tionally, so we can't consume it in the main
- * block.
- */
- memcpy(out, initial, sizeof(initial));
- num_out = initial_num_out;
- num_out_sgpr = initial_num_out_sgpr;
- continue;
- }
-
- /* Extract the returned GPRs. */
- ret_type = LLVMTypeOf(ret);
- num_out = 0;
- num_out_sgpr = 0;
-
- if (LLVMGetTypeKind(ret_type) != LLVMVoidTypeKind) {
- assert(LLVMGetTypeKind(ret_type) == LLVMStructTypeKind);
-
- unsigned ret_size = LLVMCountStructElementTypes(ret_type);
-
- for (unsigned i = 0; i < ret_size; ++i) {
- LLVMValueRef val =
- LLVMBuildExtractValue(builder, ret, i, "");
-
- assert(num_out < ARRAY_SIZE(out));
- out[num_out++] = val;
-
- if (LLVMTypeOf(val) == ctx->i32) {
- assert(num_out_sgpr + 1 == num_out);
- num_out_sgpr = num_out;
- }
- }
- }
- }
-
- /* Return the value from the last part. */
- if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
- LLVMBuildRetVoid(builder);
- else
- LLVMBuildRet(builder, ret);
+static void si_get_vs_prolog_key(const struct si_shader_info *info, unsigned num_input_sgprs,
+ bool ngg_cull_shader, const struct si_vs_prolog_bits *prolog_key,
+ struct si_shader *shader_out, union si_shader_part_key *key)
+{
+ memset(key, 0, sizeof(*key));
+ key->vs_prolog.states = *prolog_key;
+ key->vs_prolog.num_input_sgprs = num_input_sgprs;
+ key->vs_prolog.num_inputs = info->num_inputs;
+ key->vs_prolog.as_ls = shader_out->key.as_ls;
+ key->vs_prolog.as_es = shader_out->key.as_es;
+ key->vs_prolog.as_ngg = shader_out->key.as_ngg;
+ key->vs_prolog.as_prim_discard_cs = shader_out->key.opt.vs_as_prim_discard_cs;
+
+ if (ngg_cull_shader) {
+ key->vs_prolog.gs_fast_launch_tri_list =
+ !!(shader_out->key.opt.ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST);
+ key->vs_prolog.gs_fast_launch_tri_strip =
+ !!(shader_out->key.opt.ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP);
+ } else {
+ key->vs_prolog.has_ngg_cull_inputs = !!shader_out->key.opt.ngg_culling;
+ }
+
+ if (shader_out->selector->info.stage == MESA_SHADER_TESS_CTRL) {
+ key->vs_prolog.as_ls = 1;
+ key->vs_prolog.num_merged_next_stage_vgprs = 2;
+ } else if (shader_out->selector->info.stage == MESA_SHADER_GEOMETRY) {
+ key->vs_prolog.as_es = 1;
+ key->vs_prolog.num_merged_next_stage_vgprs = 5;
+ } else if (shader_out->key.as_ngg) {
+ key->vs_prolog.num_merged_next_stage_vgprs = 5;
+ }
+
+ /* Only one of these combinations can be set. as_ngg can be set with as_es. */
+ assert(key->vs_prolog.as_ls + key->vs_prolog.as_ngg +
+ (key->vs_prolog.as_es && !key->vs_prolog.as_ngg) + key->vs_prolog.as_prim_discard_cs <=
+ 1);
+
+ /* Enable loading the InstanceID VGPR. */
+ uint16_t input_mask = u_bit_consecutive(0, info->num_inputs);
+
+ if ((key->vs_prolog.states.instance_divisor_is_one |
+ key->vs_prolog.states.instance_divisor_is_fetched) &
+ input_mask)
+ shader_out->info.uses_instanceid = true;
}
static bool si_should_optimize_less(struct ac_llvm_compiler *compiler,
- struct si_shader_selector *sel)
-{
- if (!compiler->low_opt_passes)
- return false;
-
- /* Assume a slow CPU. */
- assert(!sel->screen->info.has_dedicated_vram &&
- sel->screen->info.chip_class <= GFX8);
-
- /* For a crazy dEQP test containing 2597 memory opcodes, mostly
- * buffer stores. */
- return sel->type == PIPE_SHADER_COMPUTE &&
- sel->info.num_memory_instructions > 1000;
-}
-
-static struct nir_shader *get_nir_shader(struct si_shader_selector *sel,
- bool *free_nir)
-{
- *free_nir = false;
-
- if (sel->nir) {
- return sel->nir;
- } else if (sel->nir_binary) {
- struct pipe_screen *screen = &sel->screen->b;
- const void *options =
- screen->get_compiler_options(screen, PIPE_SHADER_IR_NIR,
- sel->type);
-
- struct blob_reader blob_reader;
- blob_reader_init(&blob_reader, sel->nir_binary, sel->nir_size);
- *free_nir = true;
- return nir_deserialize(NULL, options, &blob_reader);
- }
- return NULL;
-}
-
-int si_compile_shader(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
-{
- struct si_shader_selector *sel = shader->selector;
- struct si_shader_context ctx;
- bool free_nir;
- struct nir_shader *nir = get_nir_shader(sel, &free_nir);
- int r = -1;
-
- /* Dump NIR before doing NIR->LLVM conversion in case the
- * conversion fails. */
- if (si_can_dump_shader(sscreen, sel->type) &&
- !(sscreen->debug_flags & DBG(NO_NIR))) {
- nir_print_shader(nir, stderr);
- si_dump_streamout(&sel->so);
- }
-
- si_llvm_context_init(&ctx, sscreen, compiler, si_get_shader_wave_size(shader));
- si_llvm_context_set_ir(&ctx, shader);
-
- memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
- sizeof(shader->info.vs_output_param_offset));
-
- shader->info.uses_instanceid = sel->info.uses_instanceid;
-
- if (!si_build_main_function(&ctx, nir, free_nir)) {
- si_llvm_dispose(&ctx);
- return -1;
- }
-
- if (shader->is_monolithic && ctx.type == PIPE_SHADER_VERTEX) {
- LLVMValueRef parts[2];
- bool need_prolog = si_vs_needs_prolog(sel, &shader->key.part.vs.prolog);
-
- parts[1] = ctx.main_fn;
-
- if (need_prolog) {
- union si_shader_part_key prolog_key;
- si_get_vs_prolog_key(&sel->info,
- shader->info.num_input_sgprs,
- &shader->key.part.vs.prolog,
- shader, &prolog_key);
- prolog_key.vs_prolog.is_monolithic = true;
- si_build_vs_prolog_function(&ctx, &prolog_key);
- parts[0] = ctx.main_fn;
- }
-
- si_build_wrapper_function(&ctx, parts + !need_prolog,
- 1 + need_prolog, need_prolog, 0);
-
- if (ctx.shader->key.opt.vs_as_prim_discard_cs)
- si_build_prim_discard_compute_shader(&ctx);
- } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_TESS_CTRL) {
- if (sscreen->info.chip_class >= GFX9) {
- struct si_shader_selector *ls = shader->key.part.tcs.ls;
- LLVMValueRef parts[4];
- bool vs_needs_prolog =
- si_vs_needs_prolog(ls, &shader->key.part.tcs.ls_prolog);
-
- /* TCS main part */
- parts[2] = ctx.main_fn;
-
- /* TCS epilog */
- union si_shader_part_key tcs_epilog_key;
- memset(&tcs_epilog_key, 0, sizeof(tcs_epilog_key));
- tcs_epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
- si_llvm_build_tcs_epilog(&ctx, &tcs_epilog_key);
- parts[3] = ctx.main_fn;
-
- /* VS as LS main part */
- nir = get_nir_shader(ls, &free_nir);
- struct si_shader shader_ls = {};
- shader_ls.selector = ls;
- shader_ls.key.as_ls = 1;
- shader_ls.key.mono = shader->key.mono;
- shader_ls.key.opt = shader->key.opt;
- shader_ls.is_monolithic = true;
- si_llvm_context_set_ir(&ctx, &shader_ls);
-
- if (!si_build_main_function(&ctx, nir, free_nir)) {
- si_llvm_dispose(&ctx);
- return -1;
- }
- shader->info.uses_instanceid |= ls->info.uses_instanceid;
- parts[1] = ctx.main_fn;
-
- /* LS prolog */
- if (vs_needs_prolog) {
- union si_shader_part_key vs_prolog_key;
- si_get_vs_prolog_key(&ls->info,
- shader_ls.info.num_input_sgprs,
- &shader->key.part.tcs.ls_prolog,
- shader, &vs_prolog_key);
- vs_prolog_key.vs_prolog.is_monolithic = true;
- si_build_vs_prolog_function(&ctx, &vs_prolog_key);
- parts[0] = ctx.main_fn;
- }
-
- /* Reset the shader context. */
- ctx.shader = shader;
- ctx.type = PIPE_SHADER_TESS_CTRL;
-
- si_build_wrapper_function(&ctx,
- parts + !vs_needs_prolog,
- 4 - !vs_needs_prolog, vs_needs_prolog,
- vs_needs_prolog ? 2 : 1);
- } else {
- LLVMValueRef parts[2];
- union si_shader_part_key epilog_key;
-
- parts[0] = ctx.main_fn;
-
- memset(&epilog_key, 0, sizeof(epilog_key));
- epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
- si_llvm_build_tcs_epilog(&ctx, &epilog_key);
- parts[1] = ctx.main_fn;
-
- si_build_wrapper_function(&ctx, parts, 2, 0, 0);
- }
- } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_GEOMETRY) {
- if (ctx.screen->info.chip_class >= GFX9) {
- struct si_shader_selector *es = shader->key.part.gs.es;
- LLVMValueRef es_prolog = NULL;
- LLVMValueRef es_main = NULL;
- LLVMValueRef gs_prolog = NULL;
- LLVMValueRef gs_main = ctx.main_fn;
-
- /* GS prolog */
- union si_shader_part_key gs_prolog_key;
- memset(&gs_prolog_key, 0, sizeof(gs_prolog_key));
- gs_prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
- gs_prolog_key.gs_prolog.is_monolithic = true;
- gs_prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
- si_llvm_build_gs_prolog(&ctx, &gs_prolog_key);
- gs_prolog = ctx.main_fn;
-
- /* ES main part */
- nir = get_nir_shader(es, &free_nir);
- struct si_shader shader_es = {};
- shader_es.selector = es;
- shader_es.key.as_es = 1;
- shader_es.key.as_ngg = shader->key.as_ngg;
- shader_es.key.mono = shader->key.mono;
- shader_es.key.opt = shader->key.opt;
- shader_es.is_monolithic = true;
- si_llvm_context_set_ir(&ctx, &shader_es);
-
- if (!si_build_main_function(&ctx, nir, free_nir)) {
- si_llvm_dispose(&ctx);
- return -1;
- }
- shader->info.uses_instanceid |= es->info.uses_instanceid;
- es_main = ctx.main_fn;
-
- /* ES prolog */
- if (es->type == PIPE_SHADER_VERTEX &&
- si_vs_needs_prolog(es, &shader->key.part.gs.vs_prolog)) {
- union si_shader_part_key vs_prolog_key;
- si_get_vs_prolog_key(&es->info,
- shader_es.info.num_input_sgprs,
- &shader->key.part.gs.vs_prolog,
- shader, &vs_prolog_key);
- vs_prolog_key.vs_prolog.is_monolithic = true;
- si_build_vs_prolog_function(&ctx, &vs_prolog_key);
- es_prolog = ctx.main_fn;
- }
-
- /* Reset the shader context. */
- ctx.shader = shader;
- ctx.type = PIPE_SHADER_GEOMETRY;
-
- /* Prepare the array of shader parts. */
- LLVMValueRef parts[4];
- unsigned num_parts = 0, main_part, next_first_part;
-
- if (es_prolog)
- parts[num_parts++] = es_prolog;
-
- parts[main_part = num_parts++] = es_main;
- parts[next_first_part = num_parts++] = gs_prolog;
- parts[num_parts++] = gs_main;
-
- si_build_wrapper_function(&ctx, parts, num_parts,
- main_part, next_first_part);
- } else {
- LLVMValueRef parts[2];
- union si_shader_part_key prolog_key;
-
- parts[1] = ctx.main_fn;
-
- memset(&prolog_key, 0, sizeof(prolog_key));
- prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
- si_llvm_build_gs_prolog(&ctx, &prolog_key);
- parts[0] = ctx.main_fn;
-
- si_build_wrapper_function(&ctx, parts, 2, 1, 0);
- }
- } else if (shader->is_monolithic && ctx.type == PIPE_SHADER_FRAGMENT) {
- si_llvm_build_monolithic_ps(&ctx, shader);
- }
-
- si_llvm_optimize_module(&ctx);
-
- /* Post-optimization transformations and analysis. */
- si_optimize_vs_outputs(&ctx);
-
- if ((debug && debug->debug_message) ||
- si_can_dump_shader(sscreen, ctx.type)) {
- ctx.shader->info.private_mem_vgprs =
- ac_count_scratch_private_memory(ctx.main_fn);
- }
-
- /* Make sure the input is a pointer and not integer followed by inttoptr. */
- assert(LLVMGetTypeKind(LLVMTypeOf(LLVMGetParam(ctx.main_fn, 0))) ==
- LLVMPointerTypeKind);
-
- /* Compile to bytecode. */
- r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler,
- &ctx.ac, debug, ctx.type, si_get_shader_name(shader),
- si_should_optimize_less(compiler, shader->selector));
- si_llvm_dispose(&ctx);
- if (r) {
- fprintf(stderr, "LLVM failed to compile shader\n");
- return r;
- }
-
- /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
- * LLVM 3.9svn has this bug.
- */
- if (sel->type == PIPE_SHADER_COMPUTE) {
- unsigned wave_size = sscreen->compute_wave_size;
- unsigned max_vgprs = sscreen->info.num_physical_wave64_vgprs_per_simd *
- (wave_size == 32 ? 2 : 1);
- unsigned max_sgprs = sscreen->info.num_physical_sgprs_per_simd;
- unsigned max_sgprs_per_wave = 128;
- unsigned simds_per_tg = 4; /* assuming WGP mode on gfx10 */
- unsigned threads_per_tg = si_get_max_workgroup_size(shader);
- unsigned waves_per_tg = DIV_ROUND_UP(threads_per_tg, wave_size);
- unsigned waves_per_simd = DIV_ROUND_UP(waves_per_tg, simds_per_tg);
-
- max_vgprs = max_vgprs / waves_per_simd;
- max_sgprs = MIN2(max_sgprs / waves_per_simd, max_sgprs_per_wave);
-
- if (shader->config.num_sgprs > max_sgprs ||
- shader->config.num_vgprs > max_vgprs) {
- fprintf(stderr, "LLVM failed to compile a shader correctly: "
- "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
- shader->config.num_sgprs, shader->config.num_vgprs,
- max_sgprs, max_vgprs);
-
- /* Just terminate the process, because dependent
- * shaders can hang due to bad input data, but use
- * the env var to allow shader-db to work.
- */
- if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
- abort();
- }
- }
-
- /* Add the scratch offset to input SGPRs. */
- if (shader->config.scratch_bytes_per_wave && !si_is_merged_shader(&ctx))
- shader->info.num_input_sgprs += 1; /* scratch byte offset */
-
- /* Calculate the number of fragment input VGPRs. */
- if (ctx.type == PIPE_SHADER_FRAGMENT) {
- shader->info.num_input_vgprs = ac_get_fs_input_vgpr_cnt(&shader->config,
- &shader->info.face_vgpr_index,
- &shader->info.ancillary_vgpr_index);
- }
-
- si_calculate_max_simd_waves(shader);
- si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
- return 0;
+ struct si_shader_selector *sel)
+{
+ if (!compiler->low_opt_passes)
+ return false;
+
+ /* Assume a slow CPU. */
+ assert(!sel->screen->info.has_dedicated_vram && sel->screen->info.chip_class <= GFX8);
+
+ /* For a crazy dEQP test containing 2597 memory opcodes, mostly
+ * buffer stores. */
+ return sel->info.stage == MESA_SHADER_COMPUTE && sel->info.num_memory_instructions > 1000;
+}
+
+static struct nir_shader *get_nir_shader(struct si_shader_selector *sel, bool *free_nir)
+{
+ *free_nir = false;
+
+ if (sel->nir) {
+ return sel->nir;
+ } else if (sel->nir_binary) {
+ struct pipe_screen *screen = &sel->screen->b;
+ const void *options = screen->get_compiler_options(screen, PIPE_SHADER_IR_NIR,
+ pipe_shader_type_from_mesa(sel->info.stage));
+
+ struct blob_reader blob_reader;
+ blob_reader_init(&blob_reader, sel->nir_binary, sel->nir_size);
+ *free_nir = true;
+ return nir_deserialize(NULL, options, &blob_reader);
+ }
+ return NULL;
+}
+
+static bool si_llvm_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug,
+ struct nir_shader *nir, bool free_nir)
+{
+ struct si_shader_selector *sel = shader->selector;
+ struct si_shader_context ctx;
+
+ si_llvm_context_init(&ctx, sscreen, compiler, si_get_shader_wave_size(shader));
+
+ LLVMValueRef ngg_cull_main_fn = NULL;
+ if (shader->key.opt.ngg_culling) {
+ if (!si_build_main_function(&ctx, shader, nir, false, true)) {
+ si_llvm_dispose(&ctx);
+ return false;
+ }
+ ngg_cull_main_fn = ctx.main_fn;
+ ctx.main_fn = NULL;
+ }
+
+ if (!si_build_main_function(&ctx, shader, nir, free_nir, false)) {
+ si_llvm_dispose(&ctx);
+ return false;
+ }
+
+ if (shader->is_monolithic && ctx.stage == MESA_SHADER_VERTEX) {
+ LLVMValueRef parts[4];
+ unsigned num_parts = 0;
+ bool has_prolog = false;
+ LLVMValueRef main_fn = ctx.main_fn;
+
+ if (ngg_cull_main_fn) {
+ if (si_vs_needs_prolog(sel, &shader->key.part.vs.prolog, &shader->key, true)) {
+ union si_shader_part_key prolog_key;
+ si_get_vs_prolog_key(&sel->info, shader->info.num_input_sgprs, true,
+ &shader->key.part.vs.prolog, shader, &prolog_key);
+ prolog_key.vs_prolog.is_monolithic = true;
+ si_llvm_build_vs_prolog(&ctx, &prolog_key);
+ parts[num_parts++] = ctx.main_fn;
+ has_prolog = true;
+ }
+ parts[num_parts++] = ngg_cull_main_fn;
+ }
+
+ if (si_vs_needs_prolog(sel, &shader->key.part.vs.prolog, &shader->key, false)) {
+ union si_shader_part_key prolog_key;
+ si_get_vs_prolog_key(&sel->info, shader->info.num_input_sgprs, false,
+ &shader->key.part.vs.prolog, shader, &prolog_key);
+ prolog_key.vs_prolog.is_monolithic = true;
+ si_llvm_build_vs_prolog(&ctx, &prolog_key);
+ parts[num_parts++] = ctx.main_fn;
+ has_prolog = true;
+ }
+ parts[num_parts++] = main_fn;
+
+ si_build_wrapper_function(&ctx, parts, num_parts, has_prolog ? 1 : 0, 0);
+
+ if (ctx.shader->key.opt.vs_as_prim_discard_cs)
+ si_build_prim_discard_compute_shader(&ctx);
+ } else if (shader->is_monolithic && ctx.stage == MESA_SHADER_TESS_EVAL && ngg_cull_main_fn) {
+ LLVMValueRef parts[2];
+
+ parts[0] = ngg_cull_main_fn;
+ parts[1] = ctx.main_fn;
+
+ si_build_wrapper_function(&ctx, parts, 2, 0, 0);
+ } else if (shader->is_monolithic && ctx.stage == MESA_SHADER_TESS_CTRL) {
+ if (sscreen->info.chip_class >= GFX9) {
+ struct si_shader_selector *ls = shader->key.part.tcs.ls;
+ LLVMValueRef parts[4];
+ bool vs_needs_prolog =
+ si_vs_needs_prolog(ls, &shader->key.part.tcs.ls_prolog, &shader->key, false);
+
+ /* TCS main part */
+ parts[2] = ctx.main_fn;
+
+ /* TCS epilog */
+ union si_shader_part_key tcs_epilog_key;
+ memset(&tcs_epilog_key, 0, sizeof(tcs_epilog_key));
+ tcs_epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
+ si_llvm_build_tcs_epilog(&ctx, &tcs_epilog_key);
+ parts[3] = ctx.main_fn;
+
+ /* VS as LS main part */
+ nir = get_nir_shader(ls, &free_nir);
+ struct si_shader shader_ls = {};
+ shader_ls.selector = ls;
+ shader_ls.key.as_ls = 1;
+ shader_ls.key.mono = shader->key.mono;
+ shader_ls.key.opt = shader->key.opt;
+ shader_ls.is_monolithic = true;
+
+ if (!si_build_main_function(&ctx, &shader_ls, nir, free_nir, false)) {
+ si_llvm_dispose(&ctx);
+ return false;
+ }
+ shader->info.uses_instanceid |= ls->info.uses_instanceid;
+ parts[1] = ctx.main_fn;
+
+ /* LS prolog */
+ if (vs_needs_prolog) {
+ union si_shader_part_key vs_prolog_key;
+ si_get_vs_prolog_key(&ls->info, shader_ls.info.num_input_sgprs, false,
+ &shader->key.part.tcs.ls_prolog, shader, &vs_prolog_key);
+ vs_prolog_key.vs_prolog.is_monolithic = true;
+ si_llvm_build_vs_prolog(&ctx, &vs_prolog_key);
+ parts[0] = ctx.main_fn;
+ }
+
+ /* Reset the shader context. */
+ ctx.shader = shader;
+ ctx.stage = MESA_SHADER_TESS_CTRL;
+
+ si_build_wrapper_function(&ctx, parts + !vs_needs_prolog, 4 - !vs_needs_prolog,
+ vs_needs_prolog, vs_needs_prolog ? 2 : 1);
+ } else {
+ LLVMValueRef parts[2];
+ union si_shader_part_key epilog_key;
+
+ parts[0] = ctx.main_fn;
+
+ memset(&epilog_key, 0, sizeof(epilog_key));
+ epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
+ si_llvm_build_tcs_epilog(&ctx, &epilog_key);
+ parts[1] = ctx.main_fn;
+
+ si_build_wrapper_function(&ctx, parts, 2, 0, 0);
+ }
+ } else if (shader->is_monolithic && ctx.stage == MESA_SHADER_GEOMETRY) {
+ if (ctx.screen->info.chip_class >= GFX9) {
+ struct si_shader_selector *es = shader->key.part.gs.es;
+ LLVMValueRef es_prolog = NULL;
+ LLVMValueRef es_main = NULL;
+ LLVMValueRef gs_prolog = NULL;
+ LLVMValueRef gs_main = ctx.main_fn;
+
+ /* GS prolog */
+ union si_shader_part_key gs_prolog_key;
+ memset(&gs_prolog_key, 0, sizeof(gs_prolog_key));
+ gs_prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+ gs_prolog_key.gs_prolog.is_monolithic = true;
+ gs_prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
+ si_llvm_build_gs_prolog(&ctx, &gs_prolog_key);
+ gs_prolog = ctx.main_fn;
+
+ /* ES main part */
+ nir = get_nir_shader(es, &free_nir);
+ struct si_shader shader_es = {};
+ shader_es.selector = es;
+ shader_es.key.as_es = 1;
+ shader_es.key.as_ngg = shader->key.as_ngg;
+ shader_es.key.mono = shader->key.mono;
+ shader_es.key.opt = shader->key.opt;
+ shader_es.is_monolithic = true;
+
+ if (!si_build_main_function(&ctx, &shader_es, nir, free_nir, false)) {
+ si_llvm_dispose(&ctx);
+ return false;
+ }
+ shader->info.uses_instanceid |= es->info.uses_instanceid;
+ es_main = ctx.main_fn;
+
+ /* ES prolog */
+ if (es->info.stage == MESA_SHADER_VERTEX &&
+ si_vs_needs_prolog(es, &shader->key.part.gs.vs_prolog, &shader->key, false)) {
+ union si_shader_part_key vs_prolog_key;
+ si_get_vs_prolog_key(&es->info, shader_es.info.num_input_sgprs, false,
+ &shader->key.part.gs.vs_prolog, shader, &vs_prolog_key);
+ vs_prolog_key.vs_prolog.is_monolithic = true;
+ si_llvm_build_vs_prolog(&ctx, &vs_prolog_key);
+ es_prolog = ctx.main_fn;
+ }
+
+ /* Reset the shader context. */
+ ctx.shader = shader;
+ ctx.stage = MESA_SHADER_GEOMETRY;
+
+ /* Prepare the array of shader parts. */
+ LLVMValueRef parts[4];
+ unsigned num_parts = 0, main_part, next_first_part;
+
+ if (es_prolog)
+ parts[num_parts++] = es_prolog;
+
+ parts[main_part = num_parts++] = es_main;
+ parts[next_first_part = num_parts++] = gs_prolog;
+ parts[num_parts++] = gs_main;
+
+ si_build_wrapper_function(&ctx, parts, num_parts, main_part, next_first_part);
+ } else {
+ LLVMValueRef parts[2];
+ union si_shader_part_key prolog_key;
+
+ parts[1] = ctx.main_fn;
+
+ memset(&prolog_key, 0, sizeof(prolog_key));
+ prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+ si_llvm_build_gs_prolog(&ctx, &prolog_key);
+ parts[0] = ctx.main_fn;
+
+ si_build_wrapper_function(&ctx, parts, 2, 1, 0);
+ }
+ } else if (shader->is_monolithic && ctx.stage == MESA_SHADER_FRAGMENT) {
+ si_llvm_build_monolithic_ps(&ctx, shader);
+ }
+
+ si_llvm_optimize_module(&ctx);
+
+ /* Post-optimization transformations and analysis. */
+ si_optimize_vs_outputs(&ctx);
+
+ if ((debug && debug->debug_message) || si_can_dump_shader(sscreen, ctx.stage)) {
+ ctx.shader->info.private_mem_vgprs = ac_count_scratch_private_memory(ctx.main_fn);
+ }
+
+ /* Make sure the input is a pointer and not integer followed by inttoptr. */
+ assert(LLVMGetTypeKind(LLVMTypeOf(LLVMGetParam(ctx.main_fn, 0))) == LLVMPointerTypeKind);
+
+ /* Compile to bytecode. */
+ if (!si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler, &ctx.ac, debug,
+ ctx.stage, si_get_shader_name(shader),
+ si_should_optimize_less(compiler, shader->selector))) {
+ si_llvm_dispose(&ctx);
+ fprintf(stderr, "LLVM failed to compile shader\n");
+ return false;
+ }
+
+ si_llvm_dispose(&ctx);
+ return true;
+}
+
+bool si_compile_shader(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
+{
+ struct si_shader_selector *sel = shader->selector;
+ bool free_nir;
+ struct nir_shader *nir = get_nir_shader(sel, &free_nir);
+
+ /* Dump NIR before doing NIR->LLVM conversion in case the
+ * conversion fails. */
+ if (si_can_dump_shader(sscreen, sel->info.stage) &&
+ !(sscreen->debug_flags & DBG(NO_NIR))) {
+ nir_print_shader(nir, stderr);
+ si_dump_streamout(&sel->so);
+ }
+
+ memset(shader->info.vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
+ sizeof(shader->info.vs_output_param_offset));
+
+ shader->info.uses_instanceid = sel->info.uses_instanceid;
+
+ /* TODO: ACO could compile non-monolithic shaders here (starting
+ * with PS and NGG VS), but monolithic shaders should be compiled
+ * by LLVM due to more complicated compilation.
+ */
+ if (!si_llvm_compile_shader(sscreen, compiler, shader, debug, nir, free_nir))
+ return false;
+
+ /* Validate SGPR and VGPR usage for compute to detect compiler bugs.
+ * LLVM 3.9svn has this bug.
+ */
+ if (sel->info.stage == MESA_SHADER_COMPUTE) {
+ unsigned wave_size = sscreen->compute_wave_size;
+ unsigned max_vgprs =
+ sscreen->info.num_physical_wave64_vgprs_per_simd * (wave_size == 32 ? 2 : 1);
+ unsigned max_sgprs = sscreen->info.num_physical_sgprs_per_simd;
+ unsigned max_sgprs_per_wave = 128;
+ unsigned simds_per_tg = 4; /* assuming WGP mode on gfx10 */
+ unsigned threads_per_tg = si_get_max_workgroup_size(shader);
+ unsigned waves_per_tg = DIV_ROUND_UP(threads_per_tg, wave_size);
+ unsigned waves_per_simd = DIV_ROUND_UP(waves_per_tg, simds_per_tg);
+
+ max_vgprs = max_vgprs / waves_per_simd;
+ max_sgprs = MIN2(max_sgprs / waves_per_simd, max_sgprs_per_wave);
+
+ if (shader->config.num_sgprs > max_sgprs || shader->config.num_vgprs > max_vgprs) {
+ fprintf(stderr,
+ "LLVM failed to compile a shader correctly: "
+ "SGPR:VGPR usage is %u:%u, but the hw limit is %u:%u\n",
+ shader->config.num_sgprs, shader->config.num_vgprs, max_sgprs, max_vgprs);
+
+ /* Just terminate the process, because dependent
+ * shaders can hang due to bad input data, but use
+ * the env var to allow shader-db to work.
+ */
+ if (!debug_get_bool_option("SI_PASS_BAD_SHADERS", false))
+ abort();
+ }
+ }
+
+ /* Add the scratch offset to input SGPRs. */
+ if (shader->config.scratch_bytes_per_wave && !si_is_merged_shader(shader))
+ shader->info.num_input_sgprs += 1; /* scratch byte offset */
+
+ /* Calculate the number of fragment input VGPRs. */
+ if (sel->info.stage == MESA_SHADER_FRAGMENT) {
+ shader->info.num_input_vgprs = ac_get_fs_input_vgpr_cnt(
+ &shader->config, &shader->info.face_vgpr_index, &shader->info.ancillary_vgpr_index);
+ }
+
+ si_calculate_max_simd_waves(shader);
+ si_shader_dump_stats_for_shader_db(sscreen, shader, debug);
+ return true;
}
/**
* \return non-NULL on success
*/
static struct si_shader_part *
-si_get_shader_part(struct si_screen *sscreen,
- struct si_shader_part **list,
- enum pipe_shader_type type,
- bool prolog,
- union si_shader_part_key *key,
- struct ac_llvm_compiler *compiler,
- struct pipe_debug_callback *debug,
- void (*build)(struct si_shader_context *,
- union si_shader_part_key *),
- const char *name)
-{
- struct si_shader_part *result;
-
- simple_mtx_lock(&sscreen->shader_parts_mutex);
-
- /* Find existing. */
- for (result = *list; result; result = result->next) {
- if (memcmp(&result->key, key, sizeof(*key)) == 0) {
- simple_mtx_unlock(&sscreen->shader_parts_mutex);
- return result;
- }
- }
-
- /* Compile a new one. */
- result = CALLOC_STRUCT(si_shader_part);
- result->key = *key;
-
- struct si_shader shader = {};
-
- switch (type) {
- case PIPE_SHADER_VERTEX:
- shader.key.as_ls = key->vs_prolog.as_ls;
- shader.key.as_es = key->vs_prolog.as_es;
- shader.key.as_ngg = key->vs_prolog.as_ngg;
- break;
- case PIPE_SHADER_TESS_CTRL:
- assert(!prolog);
- shader.key.part.tcs.epilog = key->tcs_epilog.states;
- break;
- case PIPE_SHADER_GEOMETRY:
- assert(prolog);
- shader.key.as_ngg = key->gs_prolog.as_ngg;
- break;
- case PIPE_SHADER_FRAGMENT:
- if (prolog)
- shader.key.part.ps.prolog = key->ps_prolog.states;
- else
- shader.key.part.ps.epilog = key->ps_epilog.states;
- break;
- default:
- unreachable("bad shader part");
- }
-
- struct si_shader_context ctx;
- si_llvm_context_init(&ctx, sscreen, compiler,
- si_get_wave_size(sscreen, type, shader.key.as_ngg,
- shader.key.as_es));
- ctx.shader = &shader;
- ctx.type = type;
-
- build(&ctx, key);
-
- /* Compile. */
- si_llvm_optimize_module(&ctx);
-
- if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler,
- &ctx.ac, debug, ctx.type, name, false)) {
- FREE(result);
- result = NULL;
- goto out;
- }
-
- result->next = *list;
- *list = result;
+si_get_shader_part(struct si_screen *sscreen, struct si_shader_part **list,
+ gl_shader_stage stage, bool prolog, union si_shader_part_key *key,
+ struct ac_llvm_compiler *compiler, struct pipe_debug_callback *debug,
+ void (*build)(struct si_shader_context *, union si_shader_part_key *),
+ const char *name)
+{
+ struct si_shader_part *result;
+
+ simple_mtx_lock(&sscreen->shader_parts_mutex);
+
+ /* Find existing. */
+ for (result = *list; result; result = result->next) {
+ if (memcmp(&result->key, key, sizeof(*key)) == 0) {
+ simple_mtx_unlock(&sscreen->shader_parts_mutex);
+ return result;
+ }
+ }
+
+ /* Compile a new one. */
+ result = CALLOC_STRUCT(si_shader_part);
+ result->key = *key;
+
+ struct si_shader_selector sel = {};
+ sel.screen = sscreen;
+
+ struct si_shader shader = {};
+ shader.selector = &sel;
+
+ switch (stage) {
+ case MESA_SHADER_VERTEX:
+ shader.key.as_ls = key->vs_prolog.as_ls;
+ shader.key.as_es = key->vs_prolog.as_es;
+ shader.key.as_ngg = key->vs_prolog.as_ngg;
+ shader.key.opt.ngg_culling =
+ (key->vs_prolog.gs_fast_launch_tri_list ? SI_NGG_CULL_GS_FAST_LAUNCH_TRI_LIST : 0) |
+ (key->vs_prolog.gs_fast_launch_tri_strip ? SI_NGG_CULL_GS_FAST_LAUNCH_TRI_STRIP : 0);
+ shader.key.opt.vs_as_prim_discard_cs = key->vs_prolog.as_prim_discard_cs;
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ assert(!prolog);
+ shader.key.part.tcs.epilog = key->tcs_epilog.states;
+ break;
+ case MESA_SHADER_GEOMETRY:
+ assert(prolog);
+ shader.key.as_ngg = key->gs_prolog.as_ngg;
+ break;
+ case MESA_SHADER_FRAGMENT:
+ if (prolog)
+ shader.key.part.ps.prolog = key->ps_prolog.states;
+ else
+ shader.key.part.ps.epilog = key->ps_epilog.states;
+ break;
+ default:
+ unreachable("bad shader part");
+ }
+
+ struct si_shader_context ctx;
+ si_llvm_context_init(&ctx, sscreen, compiler,
+ si_get_wave_size(sscreen, stage,
+ shader.key.as_ngg, shader.key.as_es,
+ shader.key.opt.ngg_culling & SI_NGG_CULL_GS_FAST_LAUNCH_ALL,
+ shader.key.opt.vs_as_prim_discard_cs));
+ ctx.shader = &shader;
+ ctx.stage = stage;
+
+ build(&ctx, key);
+
+ /* Compile. */
+ si_llvm_optimize_module(&ctx);
+
+ if (!si_compile_llvm(sscreen, &result->binary, &result->config, compiler, &ctx.ac, debug,
+ ctx.stage, name, false)) {
+ FREE(result);
+ result = NULL;
+ goto out;
+ }
+
+ result->next = *list;
+ *list = result;
out:
- si_llvm_dispose(&ctx);
- simple_mtx_unlock(&sscreen->shader_parts_mutex);
- return result;
+ si_llvm_dispose(&ctx);
+ simple_mtx_unlock(&sscreen->shader_parts_mutex);
+ return result;
}
-/**
- * Build the vertex shader prolog function.
- *
- * The inputs are the same as VS (a lot of SGPRs and 4 VGPR system values).
- * All inputs are returned unmodified. The vertex load indices are
- * stored after them, which will be used by the API VS for fetching inputs.
- *
- * For example, the expected outputs for instance_divisors[] = {0, 1, 2} are:
- * input_v0,
- * input_v1,
- * input_v2,
- * input_v3,
- * (VertexID + BaseVertex),
- * (InstanceID + StartInstance),
- * (InstanceID / 2 + StartInstance)
- */
-static void si_build_vs_prolog_function(struct si_shader_context *ctx,
- union si_shader_part_key *key)
+static bool si_get_vs_prolog(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug,
+ struct si_shader *main_part, const struct si_vs_prolog_bits *key)
{
- LLVMTypeRef *returns;
- LLVMValueRef ret, func;
- int num_returns, i;
- unsigned first_vs_vgpr = key->vs_prolog.num_merged_next_stage_vgprs;
- unsigned num_input_vgprs = key->vs_prolog.num_merged_next_stage_vgprs + 4;
- struct ac_arg input_sgpr_param[key->vs_prolog.num_input_sgprs];
- struct ac_arg input_vgpr_param[9];
- LLVMValueRef input_vgprs[9];
- unsigned num_all_input_regs = key->vs_prolog.num_input_sgprs +
- num_input_vgprs;
- unsigned user_sgpr_base = key->vs_prolog.num_merged_next_stage_vgprs ? 8 : 0;
-
- memset(&ctx->args, 0, sizeof(ctx->args));
-
- /* 4 preloaded VGPRs + vertex load indices as prolog outputs */
- returns = alloca((num_all_input_regs + key->vs_prolog.num_inputs) *
- sizeof(LLVMTypeRef));
- num_returns = 0;
-
- /* Declare input and output SGPRs. */
- for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
- ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT,
- &input_sgpr_param[i]);
- returns[num_returns++] = ctx->i32;
- }
-
- struct ac_arg merged_wave_info = input_sgpr_param[3];
-
- /* Preloaded VGPRs (outputs must be floats) */
- for (i = 0; i < num_input_vgprs; i++) {
- ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_INT, &input_vgpr_param[i]);
- returns[num_returns++] = ctx->f32;
- }
-
- /* Vertex load indices. */
- for (i = 0; i < key->vs_prolog.num_inputs; i++)
- returns[num_returns++] = ctx->f32;
-
- /* Create the function. */
- si_llvm_create_func(ctx, "vs_prolog", returns, num_returns, 0);
- func = ctx->main_fn;
-
- for (i = 0; i < num_input_vgprs; i++) {
- input_vgprs[i] = ac_get_arg(&ctx->ac, input_vgpr_param[i]);
- }
-
- if (key->vs_prolog.num_merged_next_stage_vgprs) {
- if (!key->vs_prolog.is_monolithic)
- si_init_exec_from_input(ctx, merged_wave_info, 0);
-
- if (key->vs_prolog.as_ls &&
- ctx->screen->info.has_ls_vgpr_init_bug) {
- /* If there are no HS threads, SPI loads the LS VGPRs
- * starting at VGPR 0. Shift them back to where they
- * belong.
- */
- LLVMValueRef has_hs_threads =
- LLVMBuildICmp(ctx->ac.builder, LLVMIntNE,
- si_unpack_param(ctx, input_sgpr_param[3], 8, 8),
- ctx->i32_0, "");
-
- for (i = 4; i > 0; --i) {
- input_vgprs[i + 1] =
- LLVMBuildSelect(ctx->ac.builder, has_hs_threads,
- input_vgprs[i + 1],
- input_vgprs[i - 1], "");
- }
- }
- }
-
- unsigned vertex_id_vgpr = first_vs_vgpr;
- unsigned instance_id_vgpr =
- ctx->screen->info.chip_class >= GFX10 ?
- first_vs_vgpr + 3 :
- first_vs_vgpr + (key->vs_prolog.as_ls ? 2 : 1);
-
- ctx->abi.vertex_id = input_vgprs[vertex_id_vgpr];
- ctx->abi.instance_id = input_vgprs[instance_id_vgpr];
-
- /* InstanceID = VertexID >> 16;
- * VertexID = VertexID & 0xffff;
- */
- if (key->vs_prolog.states.unpack_instance_id_from_vertex_id) {
- ctx->abi.instance_id = LLVMBuildLShr(ctx->ac.builder, ctx->abi.vertex_id,
- LLVMConstInt(ctx->i32, 16, 0), "");
- ctx->abi.vertex_id = LLVMBuildAnd(ctx->ac.builder, ctx->abi.vertex_id,
- LLVMConstInt(ctx->i32, 0xffff, 0), "");
- }
-
- /* Copy inputs to outputs. This should be no-op, as the registers match,
- * but it will prevent the compiler from overwriting them unintentionally.
- */
- ret = ctx->return_value;
- for (i = 0; i < key->vs_prolog.num_input_sgprs; i++) {
- LLVMValueRef p = LLVMGetParam(func, i);
- ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
- }
- for (i = 0; i < num_input_vgprs; i++) {
- LLVMValueRef p = input_vgprs[i];
-
- if (i == vertex_id_vgpr)
- p = ctx->abi.vertex_id;
- else if (i == instance_id_vgpr)
- p = ctx->abi.instance_id;
-
- p = ac_to_float(&ctx->ac, p);
- ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p,
- key->vs_prolog.num_input_sgprs + i, "");
- }
-
- /* Compute vertex load indices from instance divisors. */
- LLVMValueRef instance_divisor_constbuf = NULL;
-
- if (key->vs_prolog.states.instance_divisor_is_fetched) {
- LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
- LLVMValueRef buf_index =
- LLVMConstInt(ctx->i32, SI_VS_CONST_INSTANCE_DIVISORS, 0);
- instance_divisor_constbuf =
- ac_build_load_to_sgpr(&ctx->ac, list, buf_index);
- }
-
- for (i = 0; i < key->vs_prolog.num_inputs; i++) {
- bool divisor_is_one =
- key->vs_prolog.states.instance_divisor_is_one & (1u << i);
- bool divisor_is_fetched =
- key->vs_prolog.states.instance_divisor_is_fetched & (1u << i);
- LLVMValueRef index = NULL;
-
- if (divisor_is_one) {
- index = ctx->abi.instance_id;
- } else if (divisor_is_fetched) {
- LLVMValueRef udiv_factors[4];
-
- for (unsigned j = 0; j < 4; j++) {
- udiv_factors[j] =
- si_buffer_load_const(ctx, instance_divisor_constbuf,
- LLVMConstInt(ctx->i32, i*16 + j*4, 0));
- udiv_factors[j] = ac_to_integer(&ctx->ac, udiv_factors[j]);
- }
- /* The faster NUW version doesn't work when InstanceID == UINT_MAX.
- * Such InstanceID might not be achievable in a reasonable time though.
- */
- index = ac_build_fast_udiv_nuw(&ctx->ac, ctx->abi.instance_id,
- udiv_factors[0], udiv_factors[1],
- udiv_factors[2], udiv_factors[3]);
- }
-
- if (divisor_is_one || divisor_is_fetched) {
- /* Add StartInstance. */
- index = LLVMBuildAdd(ctx->ac.builder, index,
- LLVMGetParam(ctx->main_fn, user_sgpr_base +
- SI_SGPR_START_INSTANCE), "");
- } else {
- /* VertexID + BaseVertex */
- index = LLVMBuildAdd(ctx->ac.builder,
- ctx->abi.vertex_id,
- LLVMGetParam(func, user_sgpr_base +
- SI_SGPR_BASE_VERTEX), "");
- }
-
- index = ac_to_float(&ctx->ac, index);
- ret = LLVMBuildInsertValue(ctx->ac.builder, ret, index,
- ctx->args.arg_count + i, "");
- }
-
- si_llvm_build_ret(ctx, ret);
-}
+ struct si_shader_selector *vs = main_part->selector;
-static bool si_get_vs_prolog(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug,
- struct si_shader *main_part,
- const struct si_vs_prolog_bits *key)
-{
- struct si_shader_selector *vs = main_part->selector;
-
- if (!si_vs_needs_prolog(vs, key))
- return true;
-
- /* Get the prolog. */
- union si_shader_part_key prolog_key;
- si_get_vs_prolog_key(&vs->info, main_part->info.num_input_sgprs,
- key, shader, &prolog_key);
-
- shader->prolog =
- si_get_shader_part(sscreen, &sscreen->vs_prologs,
- PIPE_SHADER_VERTEX, true, &prolog_key, compiler,
- debug, si_build_vs_prolog_function,
- "Vertex Shader Prolog");
- return shader->prolog != NULL;
+ if (!si_vs_needs_prolog(vs, key, &shader->key, false))
+ return true;
+
+ /* Get the prolog. */
+ union si_shader_part_key prolog_key;
+ si_get_vs_prolog_key(&vs->info, main_part->info.num_input_sgprs, false, key, shader,
+ &prolog_key);
+
+ shader->prolog =
+ si_get_shader_part(sscreen, &sscreen->vs_prologs, MESA_SHADER_VERTEX, true, &prolog_key,
+ compiler, debug, si_llvm_build_vs_prolog, "Vertex Shader Prolog");
+ return shader->prolog != NULL;
}
/**
* Select and compile (or reuse) vertex shader parts (prolog & epilog).
*/
-static bool si_shader_select_vs_parts(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
+static bool si_shader_select_vs_parts(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
{
- return si_get_vs_prolog(sscreen, compiler, shader, debug, shader,
- &shader->key.part.vs.prolog);
+ return si_get_vs_prolog(sscreen, compiler, shader, debug, shader, &shader->key.part.vs.prolog);
}
/**
* Select and compile (or reuse) TCS parts (epilog).
*/
-static bool si_shader_select_tcs_parts(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
+static bool si_shader_select_tcs_parts(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
{
- if (sscreen->info.chip_class >= GFX9) {
- struct si_shader *ls_main_part =
- shader->key.part.tcs.ls->main_shader_part_ls;
-
- if (!si_get_vs_prolog(sscreen, compiler, shader, debug, ls_main_part,
- &shader->key.part.tcs.ls_prolog))
- return false;
-
- shader->previous_stage = ls_main_part;
- }
-
- /* Get the epilog. */
- union si_shader_part_key epilog_key;
- memset(&epilog_key, 0, sizeof(epilog_key));
- epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
-
- shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs,
- PIPE_SHADER_TESS_CTRL, false,
- &epilog_key, compiler, debug,
- si_llvm_build_tcs_epilog,
- "Tessellation Control Shader Epilog");
- return shader->epilog != NULL;
+ if (sscreen->info.chip_class >= GFX9) {
+ struct si_shader *ls_main_part = shader->key.part.tcs.ls->main_shader_part_ls;
+
+ if (!si_get_vs_prolog(sscreen, compiler, shader, debug, ls_main_part,
+ &shader->key.part.tcs.ls_prolog))
+ return false;
+
+ shader->previous_stage = ls_main_part;
+ }
+
+ /* Get the epilog. */
+ union si_shader_part_key epilog_key;
+ memset(&epilog_key, 0, sizeof(epilog_key));
+ epilog_key.tcs_epilog.states = shader->key.part.tcs.epilog;
+
+ shader->epilog = si_get_shader_part(sscreen, &sscreen->tcs_epilogs, MESA_SHADER_TESS_CTRL, false,
+ &epilog_key, compiler, debug, si_llvm_build_tcs_epilog,
+ "Tessellation Control Shader Epilog");
+ return shader->epilog != NULL;
}
/**
* Select and compile (or reuse) GS parts (prolog).
*/
-static bool si_shader_select_gs_parts(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
+static bool si_shader_select_gs_parts(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
{
- if (sscreen->info.chip_class >= GFX9) {
- struct si_shader *es_main_part;
- enum pipe_shader_type es_type = shader->key.part.gs.es->type;
-
- if (shader->key.as_ngg)
- es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
- else
- es_main_part = shader->key.part.gs.es->main_shader_part_es;
-
- if (es_type == PIPE_SHADER_VERTEX &&
- !si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
- &shader->key.part.gs.vs_prolog))
- return false;
-
- shader->previous_stage = es_main_part;
- }
-
- if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
- return true;
-
- union si_shader_part_key prolog_key;
- memset(&prolog_key, 0, sizeof(prolog_key));
- prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
- prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
-
- shader->prolog2 = si_get_shader_part(sscreen, &sscreen->gs_prologs,
- PIPE_SHADER_GEOMETRY, true,
- &prolog_key, compiler, debug,
- si_llvm_build_gs_prolog,
- "Geometry Shader Prolog");
- return shader->prolog2 != NULL;
+ if (sscreen->info.chip_class >= GFX9) {
+ struct si_shader *es_main_part;
+
+ if (shader->key.as_ngg)
+ es_main_part = shader->key.part.gs.es->main_shader_part_ngg_es;
+ else
+ es_main_part = shader->key.part.gs.es->main_shader_part_es;
+
+ if (shader->key.part.gs.es->info.stage == MESA_SHADER_VERTEX &&
+ !si_get_vs_prolog(sscreen, compiler, shader, debug, es_main_part,
+ &shader->key.part.gs.vs_prolog))
+ return false;
+
+ shader->previous_stage = es_main_part;
+ }
+
+ if (!shader->key.part.gs.prolog.tri_strip_adj_fix)
+ return true;
+
+ union si_shader_part_key prolog_key;
+ memset(&prolog_key, 0, sizeof(prolog_key));
+ prolog_key.gs_prolog.states = shader->key.part.gs.prolog;
+ prolog_key.gs_prolog.as_ngg = shader->key.as_ngg;
+
+ shader->prolog2 =
+ si_get_shader_part(sscreen, &sscreen->gs_prologs, MESA_SHADER_GEOMETRY, true, &prolog_key,
+ compiler, debug, si_llvm_build_gs_prolog, "Geometry Shader Prolog");
+ return shader->prolog2 != NULL;
}
/**
* Compute the PS prolog key, which contains all the information needed to
* build the PS prolog function, and set related bits in shader->config.
*/
-void si_get_ps_prolog_key(struct si_shader *shader,
- union si_shader_part_key *key,
- bool separate_prolog)
-{
- struct si_shader_info *info = &shader->selector->info;
-
- memset(key, 0, sizeof(*key));
- key->ps_prolog.states = shader->key.part.ps.prolog;
- key->ps_prolog.colors_read = info->colors_read;
- key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
- key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
- key->ps_prolog.wqm = info->uses_derivatives &&
- (key->ps_prolog.colors_read ||
- key->ps_prolog.states.force_persp_sample_interp ||
- key->ps_prolog.states.force_linear_sample_interp ||
- key->ps_prolog.states.force_persp_center_interp ||
- key->ps_prolog.states.force_linear_center_interp ||
- key->ps_prolog.states.bc_optimize_for_persp ||
- key->ps_prolog.states.bc_optimize_for_linear);
- key->ps_prolog.ancillary_vgpr_index = shader->info.ancillary_vgpr_index;
-
- if (info->colors_read) {
- unsigned *color = shader->selector->color_attr_index;
-
- if (shader->key.part.ps.prolog.color_two_side) {
- /* BCOLORs are stored after the last input. */
- key->ps_prolog.num_interp_inputs = info->num_inputs;
- key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
- if (separate_prolog)
- shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
- }
-
- for (unsigned i = 0; i < 2; i++) {
- unsigned interp = info->input_interpolate[color[i]];
- unsigned location = info->input_interpolate_loc[color[i]];
-
- if (!(info->colors_read & (0xf << i*4)))
- continue;
-
- key->ps_prolog.color_attr_index[i] = color[i];
-
- if (shader->key.part.ps.prolog.flatshade_colors &&
- interp == TGSI_INTERPOLATE_COLOR)
- interp = TGSI_INTERPOLATE_CONSTANT;
-
- switch (interp) {
- case TGSI_INTERPOLATE_CONSTANT:
- key->ps_prolog.color_interp_vgpr_index[i] = -1;
- break;
- case TGSI_INTERPOLATE_PERSPECTIVE:
- case TGSI_INTERPOLATE_COLOR:
- /* Force the interpolation location for colors here. */
- if (shader->key.part.ps.prolog.force_persp_sample_interp)
- location = TGSI_INTERPOLATE_LOC_SAMPLE;
- if (shader->key.part.ps.prolog.force_persp_center_interp)
- location = TGSI_INTERPOLATE_LOC_CENTER;
-
- switch (location) {
- case TGSI_INTERPOLATE_LOC_SAMPLE:
- key->ps_prolog.color_interp_vgpr_index[i] = 0;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_SAMPLE_ENA(1);
- }
- break;
- case TGSI_INTERPOLATE_LOC_CENTER:
- key->ps_prolog.color_interp_vgpr_index[i] = 2;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_CENTER_ENA(1);
- }
- break;
- case TGSI_INTERPOLATE_LOC_CENTROID:
- key->ps_prolog.color_interp_vgpr_index[i] = 4;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_PERSP_CENTROID_ENA(1);
- }
- break;
- default:
- assert(0);
- }
- break;
- case TGSI_INTERPOLATE_LINEAR:
- /* Force the interpolation location for colors here. */
- if (shader->key.part.ps.prolog.force_linear_sample_interp)
- location = TGSI_INTERPOLATE_LOC_SAMPLE;
- if (shader->key.part.ps.prolog.force_linear_center_interp)
- location = TGSI_INTERPOLATE_LOC_CENTER;
-
- /* The VGPR assignment for non-monolithic shaders
- * works because InitialPSInputAddr is set on the
- * main shader and PERSP_PULL_MODEL is never used.
- */
- switch (location) {
- case TGSI_INTERPOLATE_LOC_SAMPLE:
- key->ps_prolog.color_interp_vgpr_index[i] =
- separate_prolog ? 6 : 9;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_SAMPLE_ENA(1);
- }
- break;
- case TGSI_INTERPOLATE_LOC_CENTER:
- key->ps_prolog.color_interp_vgpr_index[i] =
- separate_prolog ? 8 : 11;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_CENTER_ENA(1);
- }
- break;
- case TGSI_INTERPOLATE_LOC_CENTROID:
- key->ps_prolog.color_interp_vgpr_index[i] =
- separate_prolog ? 10 : 13;
- if (separate_prolog) {
- shader->config.spi_ps_input_ena |=
- S_0286CC_LINEAR_CENTROID_ENA(1);
- }
- break;
- default:
- assert(0);
- }
- break;
- default:
- assert(0);
- }
- }
- }
+void si_get_ps_prolog_key(struct si_shader *shader, union si_shader_part_key *key,
+ bool separate_prolog)
+{
+ struct si_shader_info *info = &shader->selector->info;
+
+ memset(key, 0, sizeof(*key));
+ key->ps_prolog.states = shader->key.part.ps.prolog;
+ key->ps_prolog.colors_read = info->colors_read;
+ key->ps_prolog.num_input_sgprs = shader->info.num_input_sgprs;
+ key->ps_prolog.num_input_vgprs = shader->info.num_input_vgprs;
+ key->ps_prolog.wqm =
+ info->uses_derivatives &&
+ (key->ps_prolog.colors_read || key->ps_prolog.states.force_persp_sample_interp ||
+ key->ps_prolog.states.force_linear_sample_interp ||
+ key->ps_prolog.states.force_persp_center_interp ||
+ key->ps_prolog.states.force_linear_center_interp ||
+ key->ps_prolog.states.bc_optimize_for_persp || key->ps_prolog.states.bc_optimize_for_linear);
+ key->ps_prolog.ancillary_vgpr_index = shader->info.ancillary_vgpr_index;
+
+ if (info->colors_read) {
+ unsigned *color = shader->selector->color_attr_index;
+
+ if (shader->key.part.ps.prolog.color_two_side) {
+ /* BCOLORs are stored after the last input. */
+ key->ps_prolog.num_interp_inputs = info->num_inputs;
+ key->ps_prolog.face_vgpr_index = shader->info.face_vgpr_index;
+ if (separate_prolog)
+ shader->config.spi_ps_input_ena |= S_0286CC_FRONT_FACE_ENA(1);
+ }
+
+ for (unsigned i = 0; i < 2; i++) {
+ unsigned interp = info->color_interpolate[i];
+ unsigned location = info->color_interpolate_loc[i];
+
+ if (!(info->colors_read & (0xf << i * 4)))
+ continue;
+
+ key->ps_prolog.color_attr_index[i] = color[i];
+
+ if (shader->key.part.ps.prolog.flatshade_colors && interp == INTERP_MODE_COLOR)
+ interp = INTERP_MODE_FLAT;
+
+ switch (interp) {
+ case INTERP_MODE_FLAT:
+ key->ps_prolog.color_interp_vgpr_index[i] = -1;
+ break;
+ case INTERP_MODE_SMOOTH:
+ case INTERP_MODE_COLOR:
+ /* Force the interpolation location for colors here. */
+ if (shader->key.part.ps.prolog.force_persp_sample_interp)
+ location = TGSI_INTERPOLATE_LOC_SAMPLE;
+ if (shader->key.part.ps.prolog.force_persp_center_interp)
+ location = TGSI_INTERPOLATE_LOC_CENTER;
+
+ switch (location) {
+ case TGSI_INTERPOLATE_LOC_SAMPLE:
+ key->ps_prolog.color_interp_vgpr_index[i] = 0;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
+ }
+ break;
+ case TGSI_INTERPOLATE_LOC_CENTER:
+ key->ps_prolog.color_interp_vgpr_index[i] = 2;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
+ }
+ break;
+ case TGSI_INTERPOLATE_LOC_CENTROID:
+ key->ps_prolog.color_interp_vgpr_index[i] = 4;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTROID_ENA(1);
+ }
+ break;
+ default:
+ assert(0);
+ }
+ break;
+ case INTERP_MODE_NOPERSPECTIVE:
+ /* Force the interpolation location for colors here. */
+ if (shader->key.part.ps.prolog.force_linear_sample_interp)
+ location = TGSI_INTERPOLATE_LOC_SAMPLE;
+ if (shader->key.part.ps.prolog.force_linear_center_interp)
+ location = TGSI_INTERPOLATE_LOC_CENTER;
+
+ /* The VGPR assignment for non-monolithic shaders
+ * works because InitialPSInputAddr is set on the
+ * main shader and PERSP_PULL_MODEL is never used.
+ */
+ switch (location) {
+ case TGSI_INTERPOLATE_LOC_SAMPLE:
+ key->ps_prolog.color_interp_vgpr_index[i] = separate_prolog ? 6 : 9;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
+ }
+ break;
+ case TGSI_INTERPOLATE_LOC_CENTER:
+ key->ps_prolog.color_interp_vgpr_index[i] = separate_prolog ? 8 : 11;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
+ }
+ break;
+ case TGSI_INTERPOLATE_LOC_CENTROID:
+ key->ps_prolog.color_interp_vgpr_index[i] = separate_prolog ? 10 : 13;
+ if (separate_prolog) {
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTROID_ENA(1);
+ }
+ break;
+ default:
+ assert(0);
+ }
+ break;
+ default:
+ assert(0);
+ }
+ }
+ }
}
/**
*/
bool si_need_ps_prolog(const union si_shader_part_key *key)
{
- return key->ps_prolog.colors_read ||
- key->ps_prolog.states.force_persp_sample_interp ||
- key->ps_prolog.states.force_linear_sample_interp ||
- key->ps_prolog.states.force_persp_center_interp ||
- key->ps_prolog.states.force_linear_center_interp ||
- key->ps_prolog.states.bc_optimize_for_persp ||
- key->ps_prolog.states.bc_optimize_for_linear ||
- key->ps_prolog.states.poly_stipple ||
- key->ps_prolog.states.samplemask_log_ps_iter;
+ return key->ps_prolog.colors_read || key->ps_prolog.states.force_persp_sample_interp ||
+ key->ps_prolog.states.force_linear_sample_interp ||
+ key->ps_prolog.states.force_persp_center_interp ||
+ key->ps_prolog.states.force_linear_center_interp ||
+ key->ps_prolog.states.bc_optimize_for_persp ||
+ key->ps_prolog.states.bc_optimize_for_linear || key->ps_prolog.states.poly_stipple ||
+ key->ps_prolog.states.samplemask_log_ps_iter;
}
/**
* Compute the PS epilog key, which contains all the information needed to
* build the PS epilog function.
*/
-void si_get_ps_epilog_key(struct si_shader *shader,
- union si_shader_part_key *key)
+void si_get_ps_epilog_key(struct si_shader *shader, union si_shader_part_key *key)
{
- struct si_shader_info *info = &shader->selector->info;
- memset(key, 0, sizeof(*key));
- key->ps_epilog.colors_written = info->colors_written;
- key->ps_epilog.writes_z = info->writes_z;
- key->ps_epilog.writes_stencil = info->writes_stencil;
- key->ps_epilog.writes_samplemask = info->writes_samplemask;
- key->ps_epilog.states = shader->key.part.ps.epilog;
+ struct si_shader_info *info = &shader->selector->info;
+ memset(key, 0, sizeof(*key));
+ key->ps_epilog.colors_written = info->colors_written;
+ key->ps_epilog.writes_z = info->writes_z;
+ key->ps_epilog.writes_stencil = info->writes_stencil;
+ key->ps_epilog.writes_samplemask = info->writes_samplemask;
+ key->ps_epilog.states = shader->key.part.ps.epilog;
}
/**
* Select and compile (or reuse) pixel shader parts (prolog & epilog).
*/
-static bool si_shader_select_ps_parts(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
-{
- union si_shader_part_key prolog_key;
- union si_shader_part_key epilog_key;
-
- /* Get the prolog. */
- si_get_ps_prolog_key(shader, &prolog_key, true);
-
- /* The prolog is a no-op if these aren't set. */
- if (si_need_ps_prolog(&prolog_key)) {
- shader->prolog =
- si_get_shader_part(sscreen, &sscreen->ps_prologs,
- PIPE_SHADER_FRAGMENT, true,
- &prolog_key, compiler, debug,
- si_llvm_build_ps_prolog,
- "Fragment Shader Prolog");
- if (!shader->prolog)
- return false;
- }
-
- /* Get the epilog. */
- si_get_ps_epilog_key(shader, &epilog_key);
-
- shader->epilog =
- si_get_shader_part(sscreen, &sscreen->ps_epilogs,
- PIPE_SHADER_FRAGMENT, false,
- &epilog_key, compiler, debug,
- si_llvm_build_ps_epilog,
- "Fragment Shader Epilog");
- if (!shader->epilog)
- return false;
-
- /* Enable POS_FIXED_PT if polygon stippling is enabled. */
- if (shader->key.part.ps.prolog.poly_stipple) {
- shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
- assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
- }
-
- /* Set up the enable bits for per-sample shading if needed. */
- if (shader->key.part.ps.prolog.force_persp_sample_interp &&
- (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
- G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
- shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
- shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
- shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
- }
- if (shader->key.part.ps.prolog.force_linear_sample_interp &&
- (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
- G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
- shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
- shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
- shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
- }
- if (shader->key.part.ps.prolog.force_persp_center_interp &&
- (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
- G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
- shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
- shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
- shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
- }
- if (shader->key.part.ps.prolog.force_linear_center_interp &&
- (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
- G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
- shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
- shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
- shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
- }
-
- /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
- if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
- !(shader->config.spi_ps_input_ena & 0xf)) {
- shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
- assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
- }
-
- /* At least one pair of interpolation weights must be enabled. */
- if (!(shader->config.spi_ps_input_ena & 0x7f)) {
- shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
- assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
- }
-
- /* Samplemask fixup requires the sample ID. */
- if (shader->key.part.ps.prolog.samplemask_log_ps_iter) {
- shader->config.spi_ps_input_ena |= S_0286CC_ANCILLARY_ENA(1);
- assert(G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr));
- }
-
- /* The sample mask input is always enabled, because the API shader always
- * passes it through to the epilog. Disable it here if it's unused.
- */
- if (!shader->key.part.ps.epilog.poly_line_smoothing &&
- !shader->selector->info.reads_samplemask)
- shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
-
- return true;
-}
-
-void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
- unsigned *lds_size)
-{
- /* If tessellation is all offchip and on-chip GS isn't used, this
- * workaround is not needed.
- */
- return;
-
- /* SPI barrier management bug:
- * Make sure we have at least 4k of LDS in use to avoid the bug.
- * It applies to workgroup sizes of more than one wavefront.
- */
- if (sscreen->info.family == CHIP_BONAIRE ||
- sscreen->info.family == CHIP_KABINI)
- *lds_size = MAX2(*lds_size, 8);
+static bool si_shader_select_ps_parts(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
+{
+ union si_shader_part_key prolog_key;
+ union si_shader_part_key epilog_key;
+
+ /* Get the prolog. */
+ si_get_ps_prolog_key(shader, &prolog_key, true);
+
+ /* The prolog is a no-op if these aren't set. */
+ if (si_need_ps_prolog(&prolog_key)) {
+ shader->prolog =
+ si_get_shader_part(sscreen, &sscreen->ps_prologs, MESA_SHADER_FRAGMENT, true, &prolog_key,
+ compiler, debug, si_llvm_build_ps_prolog, "Fragment Shader Prolog");
+ if (!shader->prolog)
+ return false;
+ }
+
+ /* Get the epilog. */
+ si_get_ps_epilog_key(shader, &epilog_key);
+
+ shader->epilog =
+ si_get_shader_part(sscreen, &sscreen->ps_epilogs, MESA_SHADER_FRAGMENT, false, &epilog_key,
+ compiler, debug, si_llvm_build_ps_epilog, "Fragment Shader Epilog");
+ if (!shader->epilog)
+ return false;
+
+ /* Enable POS_FIXED_PT if polygon stippling is enabled. */
+ if (shader->key.part.ps.prolog.poly_stipple) {
+ shader->config.spi_ps_input_ena |= S_0286CC_POS_FIXED_PT_ENA(1);
+ assert(G_0286CC_POS_FIXED_PT_ENA(shader->config.spi_ps_input_addr));
+ }
+
+ /* Set up the enable bits for per-sample shading if needed. */
+ if (shader->key.part.ps.prolog.force_persp_sample_interp &&
+ (G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_ena) ||
+ G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
+ shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTER_ENA;
+ shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_SAMPLE_ENA(1);
+ }
+ if (shader->key.part.ps.prolog.force_linear_sample_interp &&
+ (G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_ena) ||
+ G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
+ shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTER_ENA;
+ shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_SAMPLE_ENA(1);
+ }
+ if (shader->key.part.ps.prolog.force_persp_center_interp &&
+ (G_0286CC_PERSP_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
+ G_0286CC_PERSP_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
+ shader->config.spi_ps_input_ena &= C_0286CC_PERSP_SAMPLE_ENA;
+ shader->config.spi_ps_input_ena &= C_0286CC_PERSP_CENTROID_ENA;
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
+ }
+ if (shader->key.part.ps.prolog.force_linear_center_interp &&
+ (G_0286CC_LINEAR_SAMPLE_ENA(shader->config.spi_ps_input_ena) ||
+ G_0286CC_LINEAR_CENTROID_ENA(shader->config.spi_ps_input_ena))) {
+ shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_SAMPLE_ENA;
+ shader->config.spi_ps_input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
+ }
+
+ /* POW_W_FLOAT requires that one of the perspective weights is enabled. */
+ if (G_0286CC_POS_W_FLOAT_ENA(shader->config.spi_ps_input_ena) &&
+ !(shader->config.spi_ps_input_ena & 0xf)) {
+ shader->config.spi_ps_input_ena |= S_0286CC_PERSP_CENTER_ENA(1);
+ assert(G_0286CC_PERSP_CENTER_ENA(shader->config.spi_ps_input_addr));
+ }
+
+ /* At least one pair of interpolation weights must be enabled. */
+ if (!(shader->config.spi_ps_input_ena & 0x7f)) {
+ shader->config.spi_ps_input_ena |= S_0286CC_LINEAR_CENTER_ENA(1);
+ assert(G_0286CC_LINEAR_CENTER_ENA(shader->config.spi_ps_input_addr));
+ }
+
+ /* Samplemask fixup requires the sample ID. */
+ if (shader->key.part.ps.prolog.samplemask_log_ps_iter) {
+ shader->config.spi_ps_input_ena |= S_0286CC_ANCILLARY_ENA(1);
+ assert(G_0286CC_ANCILLARY_ENA(shader->config.spi_ps_input_addr));
+ }
+
+ /* The sample mask input is always enabled, because the API shader always
+ * passes it through to the epilog. Disable it here if it's unused.
+ */
+ if (!shader->key.part.ps.epilog.poly_line_smoothing && !shader->selector->info.reads_samplemask)
+ shader->config.spi_ps_input_ena &= C_0286CC_SAMPLE_COVERAGE_ENA;
+
+ return true;
+}
+
+void si_multiwave_lds_size_workaround(struct si_screen *sscreen, unsigned *lds_size)
+{
+ /* If tessellation is all offchip and on-chip GS isn't used, this
+ * workaround is not needed.
+ */
+ return;
+
+ /* SPI barrier management bug:
+ * Make sure we have at least 4k of LDS in use to avoid the bug.
+ * It applies to workgroup sizes of more than one wavefront.
+ */
+ if (sscreen->info.family == CHIP_BONAIRE || sscreen->info.family == CHIP_KABINI)
+ *lds_size = MAX2(*lds_size, 8);
}
void si_fix_resource_usage(struct si_screen *sscreen, struct si_shader *shader)
{
- unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
-
- shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
-
- if (shader->selector->type == PIPE_SHADER_COMPUTE &&
- si_get_max_workgroup_size(shader) > sscreen->compute_wave_size) {
- si_multiwave_lds_size_workaround(sscreen,
- &shader->config.lds_size);
- }
-}
-
-bool si_create_shader_variant(struct si_screen *sscreen,
- struct ac_llvm_compiler *compiler,
- struct si_shader *shader,
- struct pipe_debug_callback *debug)
-{
- struct si_shader_selector *sel = shader->selector;
- struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
- int r;
-
- /* LS, ES, VS are compiled on demand if the main part hasn't been
- * compiled for that stage.
- *
- * GS are compiled on demand if the main part hasn't been compiled
- * for the chosen NGG-ness.
- *
- * Vertex shaders are compiled on demand when a vertex fetch
- * workaround must be applied.
- */
- if (shader->is_monolithic) {
- /* Monolithic shader (compiled as a whole, has many variants,
- * may take a long time to compile).
- */
- r = si_compile_shader(sscreen, compiler, shader, debug);
- if (r)
- return false;
- } else {
- /* The shader consists of several parts:
- *
- * - the middle part is the user shader, it has 1 variant only
- * and it was compiled during the creation of the shader
- * selector
- * - the prolog part is inserted at the beginning
- * - the epilog part is inserted at the end
- *
- * The prolog and epilog have many (but simple) variants.
- *
- * Starting with gfx9, geometry and tessellation control
- * shaders also contain the prolog and user shader parts of
- * the previous shader stage.
- */
-
- if (!mainp)
- return false;
-
- /* Copy the compiled shader data over. */
- shader->is_binary_shared = true;
- shader->binary = mainp->binary;
- shader->config = mainp->config;
- shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
- shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
- shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
- shader->info.ancillary_vgpr_index = mainp->info.ancillary_vgpr_index;
- memcpy(shader->info.vs_output_param_offset,
- mainp->info.vs_output_param_offset,
- sizeof(mainp->info.vs_output_param_offset));
- shader->info.uses_instanceid = mainp->info.uses_instanceid;
- shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
- shader->info.nr_param_exports = mainp->info.nr_param_exports;
-
- /* Select prologs and/or epilogs. */
- switch (sel->type) {
- case PIPE_SHADER_VERTEX:
- if (!si_shader_select_vs_parts(sscreen, compiler, shader, debug))
- return false;
- break;
- case PIPE_SHADER_TESS_CTRL:
- if (!si_shader_select_tcs_parts(sscreen, compiler, shader, debug))
- return false;
- break;
- case PIPE_SHADER_TESS_EVAL:
- break;
- case PIPE_SHADER_GEOMETRY:
- if (!si_shader_select_gs_parts(sscreen, compiler, shader, debug))
- return false;
- break;
- case PIPE_SHADER_FRAGMENT:
- if (!si_shader_select_ps_parts(sscreen, compiler, shader, debug))
- return false;
-
- /* Make sure we have at least as many VGPRs as there
- * are allocated inputs.
- */
- shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
- shader->info.num_input_vgprs);
- break;
- default:;
- }
-
- /* Update SGPR and VGPR counts. */
- if (shader->prolog) {
- shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
- shader->prolog->config.num_sgprs);
- shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
- shader->prolog->config.num_vgprs);
- }
- if (shader->previous_stage) {
- shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
- shader->previous_stage->config.num_sgprs);
- shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
- shader->previous_stage->config.num_vgprs);
- shader->config.spilled_sgprs =
- MAX2(shader->config.spilled_sgprs,
- shader->previous_stage->config.spilled_sgprs);
- shader->config.spilled_vgprs =
- MAX2(shader->config.spilled_vgprs,
- shader->previous_stage->config.spilled_vgprs);
- shader->info.private_mem_vgprs =
- MAX2(shader->info.private_mem_vgprs,
- shader->previous_stage->info.private_mem_vgprs);
- shader->config.scratch_bytes_per_wave =
- MAX2(shader->config.scratch_bytes_per_wave,
- shader->previous_stage->config.scratch_bytes_per_wave);
- shader->info.uses_instanceid |=
- shader->previous_stage->info.uses_instanceid;
- }
- if (shader->prolog2) {
- shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
- shader->prolog2->config.num_sgprs);
- shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
- shader->prolog2->config.num_vgprs);
- }
- if (shader->epilog) {
- shader->config.num_sgprs = MAX2(shader->config.num_sgprs,
- shader->epilog->config.num_sgprs);
- shader->config.num_vgprs = MAX2(shader->config.num_vgprs,
- shader->epilog->config.num_vgprs);
- }
- si_calculate_max_simd_waves(shader);
- }
-
- if (shader->key.as_ngg) {
- assert(!shader->key.as_es && !shader->key.as_ls);
- gfx10_ngg_calculate_subgroup_info(shader);
- } else if (sscreen->info.chip_class >= GFX9 && sel->type == PIPE_SHADER_GEOMETRY) {
- gfx9_get_gs_info(shader->previous_stage_sel, sel, &shader->gs_info);
- }
-
- si_fix_resource_usage(sscreen, shader);
- si_shader_dump(sscreen, shader, debug, stderr, true);
-
- /* Upload. */
- if (!si_shader_binary_upload(sscreen, shader, 0)) {
- fprintf(stderr, "LLVM failed to upload shader\n");
- return false;
- }
-
- return true;
+ unsigned min_sgprs = shader->info.num_input_sgprs + 2; /* VCC */
+
+ shader->config.num_sgprs = MAX2(shader->config.num_sgprs, min_sgprs);
+
+ if (shader->selector->info.stage == MESA_SHADER_COMPUTE &&
+ si_get_max_workgroup_size(shader) > sscreen->compute_wave_size) {
+ si_multiwave_lds_size_workaround(sscreen, &shader->config.lds_size);
+ }
+}
+
+bool si_create_shader_variant(struct si_screen *sscreen, struct ac_llvm_compiler *compiler,
+ struct si_shader *shader, struct pipe_debug_callback *debug)
+{
+ struct si_shader_selector *sel = shader->selector;
+ struct si_shader *mainp = *si_get_main_shader_part(sel, &shader->key);
+
+ /* LS, ES, VS are compiled on demand if the main part hasn't been
+ * compiled for that stage.
+ *
+ * GS are compiled on demand if the main part hasn't been compiled
+ * for the chosen NGG-ness.
+ *
+ * Vertex shaders are compiled on demand when a vertex fetch
+ * workaround must be applied.
+ */
+ if (shader->is_monolithic) {
+ /* Monolithic shader (compiled as a whole, has many variants,
+ * may take a long time to compile).
+ */
+ if (!si_compile_shader(sscreen, compiler, shader, debug))
+ return false;
+ } else {
+ /* The shader consists of several parts:
+ *
+ * - the middle part is the user shader, it has 1 variant only
+ * and it was compiled during the creation of the shader
+ * selector
+ * - the prolog part is inserted at the beginning
+ * - the epilog part is inserted at the end
+ *
+ * The prolog and epilog have many (but simple) variants.
+ *
+ * Starting with gfx9, geometry and tessellation control
+ * shaders also contain the prolog and user shader parts of
+ * the previous shader stage.
+ */
+
+ if (!mainp)
+ return false;
+
+ /* Copy the compiled shader data over. */
+ shader->is_binary_shared = true;
+ shader->binary = mainp->binary;
+ shader->config = mainp->config;
+ shader->info.num_input_sgprs = mainp->info.num_input_sgprs;
+ shader->info.num_input_vgprs = mainp->info.num_input_vgprs;
+ shader->info.face_vgpr_index = mainp->info.face_vgpr_index;
+ shader->info.ancillary_vgpr_index = mainp->info.ancillary_vgpr_index;
+ memcpy(shader->info.vs_output_param_offset, mainp->info.vs_output_param_offset,
+ sizeof(mainp->info.vs_output_param_offset));
+ shader->info.uses_instanceid = mainp->info.uses_instanceid;
+ shader->info.nr_pos_exports = mainp->info.nr_pos_exports;
+ shader->info.nr_param_exports = mainp->info.nr_param_exports;
+
+ /* Select prologs and/or epilogs. */
+ switch (sel->info.stage) {
+ case MESA_SHADER_VERTEX:
+ if (!si_shader_select_vs_parts(sscreen, compiler, shader, debug))
+ return false;
+ break;
+ case MESA_SHADER_TESS_CTRL:
+ if (!si_shader_select_tcs_parts(sscreen, compiler, shader, debug))
+ return false;
+ break;
+ case MESA_SHADER_TESS_EVAL:
+ break;
+ case MESA_SHADER_GEOMETRY:
+ if (!si_shader_select_gs_parts(sscreen, compiler, shader, debug))
+ return false;
+ break;
+ case MESA_SHADER_FRAGMENT:
+ if (!si_shader_select_ps_parts(sscreen, compiler, shader, debug))
+ return false;
+
+ /* Make sure we have at least as many VGPRs as there
+ * are allocated inputs.
+ */
+ shader->config.num_vgprs = MAX2(shader->config.num_vgprs, shader->info.num_input_vgprs);
+ break;
+ default:;
+ }
+
+ /* Update SGPR and VGPR counts. */
+ if (shader->prolog) {
+ shader->config.num_sgprs =
+ MAX2(shader->config.num_sgprs, shader->prolog->config.num_sgprs);
+ shader->config.num_vgprs =
+ MAX2(shader->config.num_vgprs, shader->prolog->config.num_vgprs);
+ }
+ if (shader->previous_stage) {
+ shader->config.num_sgprs =
+ MAX2(shader->config.num_sgprs, shader->previous_stage->config.num_sgprs);
+ shader->config.num_vgprs =
+ MAX2(shader->config.num_vgprs, shader->previous_stage->config.num_vgprs);
+ shader->config.spilled_sgprs =
+ MAX2(shader->config.spilled_sgprs, shader->previous_stage->config.spilled_sgprs);
+ shader->config.spilled_vgprs =
+ MAX2(shader->config.spilled_vgprs, shader->previous_stage->config.spilled_vgprs);
+ shader->info.private_mem_vgprs =
+ MAX2(shader->info.private_mem_vgprs, shader->previous_stage->info.private_mem_vgprs);
+ shader->config.scratch_bytes_per_wave =
+ MAX2(shader->config.scratch_bytes_per_wave,
+ shader->previous_stage->config.scratch_bytes_per_wave);
+ shader->info.uses_instanceid |= shader->previous_stage->info.uses_instanceid;
+ }
+ if (shader->prolog2) {
+ shader->config.num_sgprs =
+ MAX2(shader->config.num_sgprs, shader->prolog2->config.num_sgprs);
+ shader->config.num_vgprs =
+ MAX2(shader->config.num_vgprs, shader->prolog2->config.num_vgprs);
+ }
+ if (shader->epilog) {
+ shader->config.num_sgprs =
+ MAX2(shader->config.num_sgprs, shader->epilog->config.num_sgprs);
+ shader->config.num_vgprs =
+ MAX2(shader->config.num_vgprs, shader->epilog->config.num_vgprs);
+ }
+ si_calculate_max_simd_waves(shader);
+ }
+
+ if (shader->key.as_ngg) {
+ assert(!shader->key.as_es && !shader->key.as_ls);
+ if (!gfx10_ngg_calculate_subgroup_info(shader)) {
+ fprintf(stderr, "Failed to compute subgroup info\n");
+ return false;
+ }
+ } else if (sscreen->info.chip_class >= GFX9 && sel->info.stage == MESA_SHADER_GEOMETRY) {
+ gfx9_get_gs_info(shader->previous_stage_sel, sel, &shader->gs_info);
+ }
+
+ si_fix_resource_usage(sscreen, shader);
+ si_shader_dump(sscreen, shader, debug, stderr, true);
+
+ /* Upload. */
+ if (!si_shader_binary_upload(sscreen, shader, 0)) {
+ fprintf(stderr, "LLVM failed to upload shader\n");
+ return false;
+ }
+
+ return true;
+}
+
+void si_shader_binary_clean(struct si_shader_binary *binary)
+{
+ free((void *)binary->elf_buffer);
+ binary->elf_buffer = NULL;
+
+ free(binary->llvm_ir_string);
+ binary->llvm_ir_string = NULL;
}
void si_shader_destroy(struct si_shader *shader)
{
- if (shader->scratch_bo)
- si_resource_reference(&shader->scratch_bo, NULL);
+ if (shader->scratch_bo)
+ si_resource_reference(&shader->scratch_bo, NULL);
- si_resource_reference(&shader->bo, NULL);
+ si_resource_reference(&shader->bo, NULL);
- if (!shader->is_binary_shared)
- si_shader_binary_clean(&shader->binary);
+ if (!shader->is_binary_shared)
+ si_shader_binary_clean(&shader->binary);
- free(shader->shader_log);
+ free(shader->shader_log);
}