radv,aco: allow unaligned LDS access on GFX9+

fossil-db (GFX10.3):
Totals from 223 (0.16% of 139391) affected shaders:
SGPRs: 10032 -> 10096 (+0.64%)
VGPRs: 7480 -> 7592 (+1.50%)
CodeSize: 853960 -> 821920 (-3.75%); split: -3.76%, +0.01%
MaxWaves: 5916 -> 5908 (-0.14%)
Instrs: 154935 -> 150281 (-3.00%); split: -3.01%, +0.01%
Cycles: 3202496 -> 3080680 (-3.80%); split: -3.81%, +0.00%
VMEM: 48187 -> 46671 (-3.15%); split: +0.29%, -3.44%
SMEM: 13869 -> 13850 (-0.14%); split: +1.52%, -1.66%
VClause: 3110 -> 3085 (-0.80%); split: -1.03%, +0.23%
SClause: 4376 -> 4381 (+0.11%)
Copies: 12132 -> 12065 (-0.55%); split: -2.61%, +2.06%
Branches: 5204 -> 5203 (-0.02%)
PreVGPRs: 6304 -> 6359 (+0.87%); split: -0.10%, +0.97%

See https://reviews.llvm.org/D82788

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8762>

src/amd/compiler/aco_instruction_selection.cpp

@@ -3466,31 +3466,35 @@ Temp lds_load_callback(Builder& bld, const LoadEmitInfo &info,
    bool large_ds_read = bld.program->chip_class >= GFX7;
    bool usable_read2 = bld.program->chip_class >= GFX7;
 
+   bool aligned2 = align % 2 == 0;
+   bool aligned4 = align % 4 == 0;
+   bool aligned8 = bld.program->dev.has_unaligned_lds_access ? aligned4 : (align % 8 == 0);
+   bool aligned16 = bld.program->dev.has_unaligned_lds_access ? aligned4 : (align % 16 == 0);
+
    bool read2 = false;
    unsigned size = 0;
    aco_opcode op;
-   //TODO: use ds_read_u8_d16_hi/ds_read_u16_d16_hi if beneficial
+   if (bytes_needed >= 16 && aligned16 && large_ds_read) {
-   if (bytes_needed >= 16 && align % 16 == 0 && large_ds_read) {
       size = 16;
       op = aco_opcode::ds_read_b128;
-   } else if (bytes_needed >= 16 && align % 8 == 0 && const_offset % 8 == 0 && usable_read2) {
+   } else if (bytes_needed >= 16 && aligned8 && const_offset % 8 == 0 && usable_read2) {
       size = 16;
       read2 = true;
       op = aco_opcode::ds_read2_b64;
-   } else if (bytes_needed >= 12 && align % 16 == 0 && large_ds_read) {
+   } else if (bytes_needed >= 12 && aligned16 && large_ds_read) {
       size = 12;
       op = aco_opcode::ds_read_b96;
-   } else if (bytes_needed >= 8 && align % 8 == 0) {
+   } else if (bytes_needed >= 8 && aligned8) {
       size = 8;
       op = aco_opcode::ds_read_b64;
-   } else if (bytes_needed >= 8 && align % 4 == 0 && const_offset % 4 == 0) {
+   } else if (bytes_needed >= 8 && aligned4 && const_offset % 4 == 0) {
       size = 8;
       read2 = true;
       op = aco_opcode::ds_read2_b32;
-   } else if (bytes_needed >= 4 && align % 4 == 0) {
+   } else if (bytes_needed >= 4 && aligned4) {
       size = 4;
       op = aco_opcode::ds_read_b32;
-   } else if (bytes_needed >= 2 && align % 2 == 0) {
+   } else if (bytes_needed >= 2 && aligned2) {
       size = 2;
       op = aco_opcode::ds_read_u16;
    } else {
@@ -3854,8 +3858,8 @@ void store_lds(isel_context *ctx, unsigned elem_size_bytes, Temp data, uint32_t
 
       bool aligned2 = offset % 2 == 0 && align % 2 == 0;
       bool aligned4 = offset % 4 == 0 && align % 4 == 0;
-      bool aligned8 = offset % 8 == 0 && align % 8 == 0;
+      bool aligned8 = bld.program->dev.has_unaligned_lds_access ? aligned4 : (offset % 8 == 0 && align % 8 == 0);
-      bool aligned16 = offset % 16 == 0 && align % 16 == 0;
+      bool aligned16 = bld.program->dev.has_unaligned_lds_access ? aligned4 : (offset % 16 == 0 && align % 16 == 0);
 
       //TODO: use ds_write_b8_d16_hi/ds_write_b16_d16_hi if beneficial
       aco_opcode op = aco_opcode::num_opcodes;

src/amd/compiler/aco_ir.cpp

@@ -98,6 +98,10 @@ void init_program(Program *program, Stage stage, struct radv_shader_info *info,
    program->dev.lds_limit = chip_class >= GFX7 ? 65536 : 32768;
    /* apparently gfx702 also has 16-bank LDS but I can't find a family for that */
    program->dev.has_16bank_lds = family == CHIP_KABINI || family == CHIP_STONEY;
+   /* GFX10 WGP has a bug making naturally aligned access required. The LLVM
+    * subtarget feature is called "FeatureLdsMisalignedBug".
+    */
+   program->dev.has_unaligned_lds_access = chip_class >= GFX9 && !(chip_class == GFX10 && wgp_mode);
 
    program->dev.vgpr_limit = 256;
    program->dev.physical_vgprs = 256;

src/amd/compiler/aco_ir.h

@@ -1799,6 +1799,7 @@ struct DeviceInfo {
    uint16_t lds_alloc_granule;
    uint32_t lds_limit; /* in bytes */
    bool has_16bank_lds;
+   bool has_unaligned_lds_access;
    uint16_t physical_sgprs;
    uint16_t physical_vgprs;
    uint16_t vgpr_limit;

src/amd/vulkan/radv_pipeline.c

@@ -3053,6 +3053,9 @@ mem_vectorize_callback(unsigned align_mul, unsigned align_offset,
                        nir_intrinsic_instr *low, nir_intrinsic_instr *high,
                        void *data)
 {
+	struct radv_device *device = data;
+	enum chip_class chip = device->physical_device->rad_info.chip_class;
+
 	if (num_components > 4)
 		return false;
 
@@ -3081,9 +3084,9 @@ mem_vectorize_callback(unsigned align_mul, unsigned align_offset,
 		FALLTHROUGH;
 	case nir_intrinsic_load_shared:
 	case nir_intrinsic_store_shared:
-		if (bit_size * num_components == 96) /* 96 bit loads require 128 bit alignment and are split otherwise */
+		if (chip < GFX9 && bit_size * num_components == 96) /* 96 bit loads require 128 bit alignment on GFX6-8 and are split otherwise */
 			return align % 16 == 0;
-		else if (bit_size * num_components == 128) /* 128 bit loads require 64 bit alignment and are split otherwise */
+		else if (chip < GFX9 && bit_size * num_components == 128) /* 128 bit loads require 64 bit alignment on GFX6-8 and are split otherwise */
 			return align % 8 == 0;
 		else
 			return align % (bit_size == 8 ? 2 : 4) == 0;
@@ -3330,6 +3333,7 @@ VkResult radv_create_shaders(struct radv_pipeline *pipeline,
 					 nir_var_mem_push_const | nir_var_mem_shared |
 					 nir_var_mem_global,
 				.callback = mem_vectorize_callback,
+				.cb_data = device,
 				.robust_modes = 0,
 			};