fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-23 17:40:11 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

pan/mdg: Support loads and stores to scratch memory

Browse source 
Similar to shared memory load/store, except giving a different memory
type to the hardware.

Add nir->scratch_size to ctx->tls_size to allocate the memory.

Tested with the Piglit OpenCL test i32-stack-array.cl.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8264>
This commit is contained in:
Icecream95 2021-01-01 01:39:49 +13:00 • committed by Marge Bot
parent 684e5aa5b8
commit 152bc5d15e
2 changed files with 9 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
src/panfrost/midgard/midgard_compile.c
View file

@ -1502,6 +1502,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
case nir_intrinsic_load_global: case nir_intrinsic_load_global:
case nir_intrinsic_load_global_constant: case nir_intrinsic_load_global_constant:
case nir_intrinsic_load_shared: case nir_intrinsic_load_shared:
case nir_intrinsic_load_scratch:
case nir_intrinsic_load_input: case nir_intrinsic_load_input:
case nir_intrinsic_load_kernel_input: case nir_intrinsic_load_kernel_input:
case nir_intrinsic_load_interpolated_input: { case nir_intrinsic_load_interpolated_input: {
@ -1510,6 +1511,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
bool is_global = instr->intrinsic == nir_intrinsic_load_global || bool is_global = instr->intrinsic == nir_intrinsic_load_global ||
instr->intrinsic == nir_intrinsic_load_global_constant; instr->intrinsic == nir_intrinsic_load_global_constant;
bool is_shared = instr->intrinsic == nir_intrinsic_load_shared; bool is_shared = instr->intrinsic == nir_intrinsic_load_shared;
bool is_scratch = instr->intrinsic == nir_intrinsic_load_scratch;
bool is_flat = instr->intrinsic == nir_intrinsic_load_input; bool is_flat = instr->intrinsic == nir_intrinsic_load_input;
bool is_kernel = instr->intrinsic == nir_intrinsic_load_kernel_input; bool is_kernel = instr->intrinsic == nir_intrinsic_load_kernel_input;
bool is_interp = instr->intrinsic == nir_intrinsic_load_interpolated_input; bool is_interp = instr->intrinsic == nir_intrinsic_load_interpolated_input;
@ -1523,7 +1525,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
t = nir_alu_type_get_base_type(t); t = nir_alu_type_get_base_type(t);
if (!(is_ubo || is_global)) { if (!(is_ubo || is_global || is_scratch)) {
offset = nir_intrinsic_base(instr); offset = nir_intrinsic_base(instr);
} }
@ -1554,7 +1556,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
uint32_t uindex = nir_src_as_uint(index) + 1; uint32_t uindex = nir_src_as_uint(index) + 1;
emit_ubo_read(ctx, &instr->instr, reg, offset, indirect_offset, 0, uindex); emit_ubo_read(ctx, &instr->instr, reg, offset, indirect_offset, 0, uindex);
} else if (is_global || is_shared) { } else if (is_global || is_shared || is_scratch) {
unsigned seg = is_global ? LDST_GLOBAL : (is_shared ? LDST_SHARED : LDST_SCRATCH); unsigned seg = is_global ? LDST_GLOBAL : (is_shared ? LDST_SHARED : LDST_SCRATCH);
emit_global(ctx, &instr->instr, true, reg, src_offset, seg); emit_global(ctx, &instr->instr, true, reg, src_offset, seg);
} else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->is_blend) { } else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->is_blend) {
@ -1780,6 +1782,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
case nir_intrinsic_store_global: case nir_intrinsic_store_global:
case nir_intrinsic_store_shared: case nir_intrinsic_store_shared:
case nir_intrinsic_store_scratch:
reg = nir_src_index(ctx, &instr->src[0]); reg = nir_src_index(ctx, &instr->src[0]);
emit_explicit_constant(ctx, reg, reg); emit_explicit_constant(ctx, reg, reg);
@ -1788,6 +1791,8 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
seg = LDST_GLOBAL; seg = LDST_GLOBAL;
else if (instr->intrinsic == nir_intrinsic_store_shared) else if (instr->intrinsic == nir_intrinsic_store_shared)
seg = LDST_SHARED; seg = LDST_SHARED;
else
seg = LDST_SCRATCH;
emit_global(ctx, &instr->instr, false, reg, &instr->src[1], seg); emit_global(ctx, &instr->instr, false, reg, &instr->src[1], seg);
break; break;
@ -2876,6 +2881,7 @@ midgard_compile_shader_nir(void *mem_ctx, nir_shader *nir,
panfrost_nir_assign_sysvals(&ctx->sysvals, ctx, nir); panfrost_nir_assign_sysvals(&ctx->sysvals, ctx, nir);
program->sysval_count = ctx->sysvals.sysval_count; program->sysval_count = ctx->sysvals.sysval_count;
memcpy(program->sysvals, ctx->sysvals.sysvals, sizeof(ctx->sysvals.sysvals[0]) * ctx->sysvals.sysval_count); memcpy(program->sysvals, ctx->sysvals.sysvals, sizeof(ctx->sysvals.sysvals[0]) * ctx->sysvals.sysval_count);
ctx->tls_size = nir->scratch_size;
nir_foreach_function(func, nir) { nir_foreach_function(func, nir) {
if (!func->impl) if (!func->impl)

2
src/panfrost/midgard/midgard_ra.c
View file

@ -1054,7 +1054,7 @@ mir_ra(compiler_context *ctx)
/* Report spilling information. spill_count is in 128-bit slots (vec4 x /* Report spilling information. spill_count is in 128-bit slots (vec4 x
* fp32), but tls_size is in bytes, so multiply by 16 */ * fp32), but tls_size is in bytes, so multiply by 16 */
ctx->tls_size = spill_count * 16; ctx->tls_size += spill_count * 16;
install_registers(ctx, l); install_registers(ctx, l);