ac/rtld: add support for Wave32

Reviewed-by: Samuel Pitoiset <[email protected]>

8 files changed, 21 insertions, 5 deletions

diff --git a/src/amd/common/ac_binary.c b/src/amd/common/ac_binary.c
index 652ff228032..6cf05daac55 100644
--- a/src/amd/common/ac_binary.c
+++ b/src/amd/common/ac_binary.c
@@ -37,10 +37,11 @@
 
 /* Parse configuration data in .AMDGPU.config section format. */
 void ac_parse_shader_binary_config(const char *data, size_t nbytes,
+				   unsigned wave_size,
 				   bool really_needs_scratch,
 				   struct ac_shader_config *conf)
 {
-	uint32_t wavesize = 0;
+	uint32_t scratch_size = 0;
 
 	for (size_t i = 0; i < nbytes; i += 8) {
 		unsigned reg = util_le32_to_cpu(*(uint32_t*)(data + i));
@@ -51,8 +52,12 @@ void ac_parse_shader_binary_config(const char *data, size_t nbytes,
 		case R_00B228_SPI_SHADER_PGM_RSRC1_GS:
 		case R_00B848_COMPUTE_PGM_RSRC1:
 		case R_00B428_SPI_SHADER_PGM_RSRC1_HS:
+			if (wave_size == 32)
+				conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 8);
+			else
+				conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
+
 			conf->num_sgprs = MAX2(conf->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
-			conf->num_vgprs = MAX2(conf->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
 			conf->float_mode =  G_00B028_FLOAT_MODE(value);
 			conf->rsrc1 = value;
 			break;
@@ -72,7 +77,7 @@ void ac_parse_shader_binary_config(const char *data, size_t nbytes,
 		case R_0286E8_SPI_TMPRING_SIZE:
 		case R_00B860_COMPUTE_TMPRING_SIZE:
 			/* WAVESIZE is in units of 256 dwords. */
-			wavesize = value;
+			scratch_size = value;
 			break;
 		case SPILLED_SGPRS:
 			conf->spilled_sgprs = value;
@@ -99,6 +104,6 @@ void ac_parse_shader_binary_config(const char *data, size_t nbytes,
 
 	if (really_needs_scratch) {
 		/* sgprs spills aren't spilling */
-	        conf->scratch_bytes_per_wave = G_00B860_WAVESIZE(wavesize) * 256 * 4;
+	        conf->scratch_bytes_per_wave = G_00B860_WAVESIZE(scratch_size) * 256 * 4;
 	}
 }
diff --git a/src/amd/common/ac_binary.h b/src/amd/common/ac_binary.h
index 53138f11810..1276df49369 100644
--- a/src/amd/common/ac_binary.h
+++ b/src/amd/common/ac_binary.h
@@ -47,6 +47,7 @@ struct ac_shader_config {
 };
 
 void ac_parse_shader_binary_config(const char *data, size_t nbytes,
+				   unsigned wave_size,
 				   bool really_needs_scratch,
 				   struct ac_shader_config *conf);
 
diff --git a/src/amd/common/ac_rtld.c b/src/amd/common/ac_rtld.c
index 55315913234..a1bb51a8a88 100644
--- a/src/amd/common/ac_rtld.c
+++ b/src/amd/common/ac_rtld.c
@@ -262,6 +262,7 @@ bool ac_rtld_open(struct ac_rtld_binary *binary,
 
 	memset(binary, 0, sizeof(*binary));
 	memcpy(&binary->options, &i.options, sizeof(binary->options));
+	binary->wave_size = i.wave_size;
 	binary->num_parts = i.num_parts;
 	binary->parts = calloc(sizeof(*binary->parts), i.num_parts);
 	if (!binary->parts)
@@ -523,7 +524,8 @@ bool ac_rtld_read_config(struct ac_rtld_binary *binary,
 
 		/* TODO: be precise about scratch use? */
 		struct ac_shader_config c = {};
-		ac_parse_shader_binary_config(config_data, config_nbytes, true, &c);
+		ac_parse_shader_binary_config(config_data, config_nbytes,
+					      binary->wave_size, true, &c);
 
 		config->num_sgprs = MAX2(config->num_sgprs, c.num_sgprs);
 		config->num_vgprs = MAX2(config->num_vgprs, c.num_vgprs);
diff --git a/src/amd/common/ac_rtld.h b/src/amd/common/ac_rtld.h
index 3f60444f85e..55faaa011db 100644
--- a/src/amd/common/ac_rtld.h
+++ b/src/amd/common/ac_rtld.h
@@ -52,6 +52,7 @@ struct ac_rtld_options {
 /* Lightweight wrapper around underlying ELF objects. */
 struct ac_rtld_binary {
 	struct ac_rtld_options options;
+	unsigned wave_size;
 
 	/* Required buffer sizes, currently read/executable only. */
 	uint64_t rx_size;
@@ -86,6 +87,7 @@ struct ac_rtld_open_info {
 	const struct radeon_info *info;
 	struct ac_rtld_options options;
 	gl_shader_stage shader_type;
+	unsigned wave_size;
 
 	unsigned num_parts;
 	const char * const *elf_ptrs; /* in-memory ELF objects of each part */
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 17ce6d3ef54..3b00bf800b9 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -961,6 +961,7 @@ radv_shader_variant_create(struct radv_device *device,
 		struct ac_rtld_open_info open_info = {
 			.info = &device->physical_device->rad_info,
 			.shader_type = binary->stage,
+			.wave_size = 64,
 			.num_parts = 1,
 			.elf_ptrs = &elf_data,
 			.elf_sizes = &elf_size,
diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index c0b59915b1f..97f692a18dc 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -68,6 +68,7 @@ static const amd_kernel_code_t *si_compute_get_code_object(
 	if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
 			.info = &sel->screen->info,
 			.shader_type = MESA_SHADER_COMPUTE,
+			.wave_size = 64,
 			.num_parts = 1,
 			.elf_ptrs = &program->shader.binary.elf_buffer,
 			.elf_sizes = &program->shader.binary.elf_size }))
diff --git a/src/gallium/drivers/radeonsi/si_debug.c b/src/gallium/drivers/radeonsi/si_debug.c
index 7d2f492fcc8..6eddfcbe8dc 100644
--- a/src/gallium/drivers/radeonsi/si_debug.c
+++ b/src/gallium/drivers/radeonsi/si_debug.c
@@ -944,6 +944,7 @@ static void si_add_split_disasm(struct si_screen *screen,
 	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 = 64,
 			.num_parts = 1,
 			.elf_ptrs = &binary->elf_buffer,
 			.elf_sizes = &binary->elf_size }))
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 8017210ccd2..b41e1e62522 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5258,6 +5258,7 @@ static bool si_shader_binary_open(struct si_screen *screen,
 				.halt_at_entry = screen->options.halt_shaders,
 			},
 			.shader_type = tgsi_processor_to_shader_stage(sel->type),
+			.wave_size = 64,
 			.num_parts = num_parts,
 			.elf_ptrs = part_elfs,
 			.elf_sizes = part_sizes,
@@ -5352,6 +5353,7 @@ static void si_shader_dump_disassembly(struct si_screen *screen,
 	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 = 64,
 			.num_parts = 1,
 			.elf_ptrs = &binary->elf_buffer,
 			.elf_sizes = &binary->elf_size }))
@@ -5642,6 +5644,7 @@ static int si_compile_llvm(struct si_screen *sscreen,
 	if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){
 			.info = &sscreen->info,
 			.shader_type = tgsi_processor_to_shader_stage(shader_type),
+			.wave_size = 64,
 			.num_parts = 1,
 			.elf_ptrs = &binary->elf_buffer,
 			.elf_sizes = &binary->elf_size }))